enabling renewable energy powered sustainable high … · 2015. 1. 8. · enabling renewable energy...
TRANSCRIPT
ENABLING RENEWABLE ENERGY POWERED SUSTAINABLE HIGH-PERFORMANCE COMPUTING
By
CHAO LI
A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
2014
© 2014 Chao Li
This work is dedicated to my family.
4
ACKNOWLEDGMENTS
First and foremost I would like to express my sincerest gratitude to my advisor,
Dr. Tao Li, who has been a pillar of support in my graduate study. He is an inspiring
mentor who teaches me how to break through various obstacles and get things done
with intelligence, creativity, and perseverance. I could not make such a strong record of
achievements without his guidance, patience, and encouragement.
My supervisory committee guided me throughout my proposal and thesis. I want
to thank Dr. Jose A.B. Fortes, Dr. Renato Figueiredo, and Dr. Prabir Barooah, for their
valuable feedback and comments. It has been a highly enjoyable learning experience
from Dr. Jose A.B. Fortes, who taught me Autonomic Computing, and Dr. Renato
Figueiredo, who taught me Virtual Computers.
Dozens of people have helped me immensely in the IDEAL Research group.
Thanks to them, my working experience at the University of Florida is productive and
enjoyable. I want to thank Dr. Wangyuan Zhang and Dr. Xin Fu, who provided
enormous amount of help when I first joined the group. I want to thank Dr. James Poe II
and Dr. Clay Hughes, who helped with proofreading in my first year of graduate study.
Thank you to those who have spent a lot of time on the development of our 2nd and 3rd
generation system prototype: Yang Hu, Juncheng Gu, Longjun Liu, and Jingling Yuan.
Also thank you to Ruijin Zhou, Ming Liu, and Amer Qouneh, who have supported me in
many ways such as experimental setup and system trouble-shooting.
This research is generously funded by a University of Florida Alumni Fellowship
Award, a Yahoo! Key Scientific Challenges Program Award in Green Computing, and a
Facebook Graduate Fellowship Award in Computer Architecture.
5
TABLE OF CONTENTS
page
ACKNOWLEDGMENTS .................................................................................................. 4
LIST OF TABLES ............................................................................................................ 7
LIST OF FIGURES .......................................................................................................... 8
LIST OF ABBREVIATIONS ........................................................................................... 11
ABSTRACT ................................................................................................................... 12
CHAPTER
1 INTRODUCTION AND BACKGROUND ................................................................. 13
2 RELATED WORK ................................................................................................... 16
3 SERVER-LEVEL RENEWABLE POWER MANAGEMENT: RETHINKING LOAD ADAPTATION SCHEMES ...................................................................................... 19
3.1 PV Systems and Load Matching ....................................................................... 21
3.2 Managing Solar Energy Powered Multi-core System ........................................ 22
3.2.1 Multi-core Aware MPP Tracking .............................................................. 24
3.2.2 Dynamic Per-Core Load Tuning .............................................................. 26
3.3 Experimental Methodology ............................................................................... 29
3.4 Evaluation Results ............................................................................................ 32
3.4.1 Power Tracking Accuracy ........................................................................ 33
3.4.2 Fixed Power Budget ................................................................................ 34
3.4.3 Solar Energy Utilization ........................................................................... 36
3.4.4 Performance Improvement ...................................................................... 37
3.5 Summary of SolarCore Design ......................................................................... 39
4 CLUSTER LEVEL RENEWABLE POWER MANAGEMENT: UNDERSTANDING THE CONTROL OVERHEAD ................................................................................. 40
4.1 Power Management Regions ............................................................................ 41
4.2 An Overview of iSwitch Architecture ................................................................. 44
4.3 Optimizing Load Tuning Activities ..................................................................... 47
4.3.1 Lazy Power Supply Tracking ................................................................... 48
4.3.2 Dynamic Power Demand Smoothing ....................................................... 49
4.4 Experimental Methodology ............................................................................... 52
4.5 Evaluation Results ............................................................................................ 54
4.6 Summary of iSwitch Design .............................................................................. 57
6
5 DATA CENTER LEVEL GREEN POWER MANAGEMENT: TOWARDS LOAD-FOLLOWING BASED DESIGN .............................................................................. 58
5.1 The Load Following Challenge ......................................................................... 60
5.2 Distributed Generation Powered Data Centers ................................................. 61
5.2.1 Power Demand Shaping Mechanism ...................................................... 62
5.2.1 Adaptive Load Performance Scaling ....................................................... 65
5.3 Experimental Methodology ............................................................................... 67
5.4 Evaluation Results ............................................................................................ 68
5.5 Summary of PDS Control .................................................................................. 72
6 ENABLING DATA CENTER SERVERS TO SCALE OUT SUSTAINABLY ............. 73
6.1 Oasis Scale-Out Model ..................................................................................... 76
6.1.1 Modular Power Sources .......................................................................... 77
6.1.2 Distributed Integration ............................................................................. 79
6.2 Implementation of Oasis ................................................................................... 80
6.2.1 Power Control Hub .................................................................................. 80
6.2.2 Bridging Server and Power Supply .......................................................... 82
6.2.3 Dynamic Energy Source Switching.......................................................... 83
6.3 Optimized Oasis Operation (O3) ....................................................................... 85
6.3.1 Managing Battery Lifetime ....................................................................... 86
6.3.2 Managing Backup Capacity ..................................................................... 87
6.3.3 Managing Server Performance ................................................................ 87
6.4 Experimental Methodology ............................................................................... 88
6.5 Evaluation Results ............................................................................................ 90
6.5.1 Load Performance and Energy Efficiency ............................................... 91
6.5.2 Battery Service Life and Backup Capacity ............................................... 92
6.6 Cost Analysis of Oasis ...................................................................................... 95
7 SUSTAINABLE COMPUTING IN THE SMART GRID AND BIG DATA ERA .......... 98
7.1 Green Data Centers Powered by Hybrid Renewable Energy Systems ............. 99
7.1.1 Hierarchical Power Management Framework ........................................ 100
7.1.2 Multi-Source Driven Power Management .............................................. 101
7.1.3 Verification Platform and Results .......................................................... 103
7.2 Towards Sustainable Power Provisioning for In-situ Server Systems ............. 105
7.2.1 Standalone Green Server Clusters ........................................................ 107
7.2.2 System Implementation and Verification ............................................... 109
7.2.3 Analyzing In-Situ Green Computing ...................................................... 110
8 CONCLUSIONS ................................................................................................... 114
LIST OF REFERENCES ............................................................................................. 115
BIOGRAPHICAL SKETCH .......................................................................................... 123
7
LIST OF TABLES
Table page 3-1 The impact of load tuning on output power behaviors ........................................ 25
3-2 The evaluated different solar power traces ......................................................... 30
3-3 Performance levels of battery-equipped PV systems ......................................... 30
3-4 Simulated SolarCore machine configuration ...................................................... 31
3-5 The simulated multi-programmed workloads ...................................................... 32
3-6 The evaluated solar power management schemes ............................................ 32
4-1 Configurations of heterogeneous clusters .......................................................... 53
4-2 The evaluated data center workload sets ........................................................... 53
4-3 The evaluated wind power supply traces ............................................................ 54
4-4 The evaluated server cluster power management schemes .............................. 54
5-1 The evaluated HPC data center workload traces ............................................... 68
5-2 The evaluated data center power management schemes .................................. 68
5-3 Mean job turnaround time of tracking based design (ST) ................................... 69
6-1 Computing platform setup for Oasis ................................................................... 89
6-2 Evaluated cloud workloads for Oasis .................................................................. 90
7-1 Key variables collected from the inSURE system log ....................................... 112
8
LIST OF FIGURES
Figure page 3-1 Three typical PV systems ................................................................................... 20
3-2 The current-voltage and power-voltage characteristics of PV cells .................... 21
3-3 The characteristics of load matching .................................................................. 22
3-4 An overview of SolarCore power management architecture ............................... 23
3-5 The flowchart of SolarCore maximum power point tracking ............................... 25
3-6 The per-core load adaptation policies of SolarCore ............................................ 29
3-7 The per-core load adaptation demonstration of SolarCore ................................. 29
3-8 Pseudo code for SolarCore per-core load adaptation ......................................... 29
3-9 MPP tracking results for different workloads in a sunny day .............................. 33
3-10 MPP tracking results for different workloads in a cloudy day .............................. 33
3-11 The histogram of the power tracking error .......................................................... 33
3-12 Operation durations on different power budget thresholds ................................. 34
3-13 Normalized solar energy usage under different fixed power budget ................... 35
3-14 Normalized performance time product (PTP) ..................................................... 35
3-15 Energy utilization across different locations ........................................................ 37
3-16 Normalized PTP with different load scheduling methods .................................... 38
4-1 Power management overhead vs. renewable energy utilization ......................... 41
4-2 Wind power output characteristics ...................................................................... 42
4-3 The probability distribution of wind speed ........................................................... 43
4-4 Power variation in wind powered server clusters (x-axis: minutes) ..................... 43
4-5 The iSwitch load power balancing scenario ........................................................ 46
4-6 The iSwitch global (facility-level) power management ........................................ 47
4-7 The iSwitch local (rack-level) power management ............................................. 47
9
4-8 Structure of the iSwitch scheduler ...................................................................... 48
4-9 The histogram of switching frequency using round-robin virtual machine selection during load migration ........................................................................... 50
4-10 Optimization timeline of iSwitch .......................................................................... 50
4-11 The average network traffic ................................................................................ 55
4-12 The peak network traffic ..................................................................................... 55
4-13 The average execution time increase (ETI) of job requests ............................... 57
4-14 The normalized overall wind energy utilization ................................................... 57
5-1 Distributed generation (DG) powered data centers ............................................ 59
5-2 Load following scenario ...................................................................................... 59
5-3 Distributed generation powered data centers ..................................................... 61
5-4 Power demand shaping demonstration .............................................................. 63
5-5 Power demand shaping algorithm ...................................................................... 65
5-6 Average job turnaround time under varying load following interval ..................... 70
5-7 Execution time of the worst 5% jobs ................................................................... 70
5-8 Dependence on utility power .............................................................................. 71
5-9 Battery lifetime degradation comparison (with distributed UPS) ......................... 72
5-10 Battery life degradation comparison (with centralized UPS) ............................... 72
6-1 Conventional data center power architecture is not ‘scale-out friendly’ .............. 74
6-2 Oasis power provisioning architecture ................................................................ 77
6-3 Distributed energy storage system (battery) ....................................................... 78
6-4 Prototype of Oasis node ..................................................................................... 81
6-5 Power management agent ................................................................................. 83
6-6 The control flow of power supply switching ........................................................ 85
6-7 Battery charging/discharging trace ..................................................................... 85
6-8 The power control decision tree of Ozone .......................................................... 88
10
6-9 The execution latency due to server performance scaling .................................. 92
6-10 The ratio of green energy usage to overall power consumption ......................... 92
6-11 The estimated battery lifetime calculated from battery usage profile .................. 93
6-12 Battery discharging profile and power demand traces ........................................ 94
6-13 The observed average battery backup capacity in our UPS systems ................. 95
6-14 A cost breakdown of Oasis ................................................................................. 95
6-15 Cost trends of deploying Oasis for scaling out ................................................... 97
7-1 High-level system architecture for GreenWorks ............................................... 100
7-2 The GreenWorks power management hierarchy .............................................. 101
7-3 Details of our three-layer simulation platform ................................................... 103
7-4 The impact of GreenWorks on job execution time ............................................ 104
7-5 Battery lifetime estimation based on its usage log ............................................ 105
7-6 Average UPS backup time (normalized to the rated value) .............................. 105
7-7 Cumulative distribution function (CDF) for UPS autonomy time ....................... 105
7-8 In-situ server system as an ancillary to future cloud ......................................... 107
7-9 In-situ server cluster with reconfigurable battery array ..................................... 108
7-10 The energy flow scenarios for InSURE ............................................................. 109
7-11 Operating modes transition of InSURE energy buffer ....................................... 109
7-12 The structure of our verification platform .......................................................... 110
7-13 Solar power budget trace and battery voltage trace ......................................... 111
7-14 Expected optimization results from our spatio-temporal management ............. 113
11
LIST OF ABBREVIATIONS
ATS Autonomic Transfer Switch
CSC Central Switch Controller
DG/DGR Distributed Generation/Distributed Generation Resource
DVFS Dynamic Voltage and Frequency Scaling
EMM Energy Management Module
ETI Execution Time Increase
GHG Greenhouse Gas
HMI Human Machine Interface
IT Information Technology
MPPT Maximum Power Point Tracking
PCH Power Control Hub
PCPG Per-Core Power Gating
PDU Power Distribution Unit
PDC Power Demand Controller
PDS Power Demand Shaping
PLC Programmable Logic Controller
PV Photovoltaic Module
REU Renewable Energy Utilization
SAB Switch Allocation Buffer
TCO Total Cost of Ownership
UPS Un-interruptible Power Supply
VM Virtual Machine
12
Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
ENABLING RENEWABLE ENERGY POWERED SUSTAINABLE HIGH-
PERFORMANCE COMPUTING
By
Chao Li
August 2014
Chair: Tao Li Cochair: José A.B. Fortes Major: Electrical and Computer Engineering
Computer system inevitably enters the landscape of design for sustainability. As
global server power demand hits a record level of thirty gigawatts, modern data centers
must develop efficient power provisioning and management schemes to alleviate their
escalating energy needs, reduce the skyrocketing power expenditures, and resolve the
looming environmental problems. To maintain the performance scaling trend of
emerging workload applications, and to eventually realize carbon-free computing, this
dissertation extends the benefits of renewable power supply to the computing system
design area. The main obstacle of coordinating renewable power supply and server
power demand is the lack of cross-layer power management solutions that span
software, hardware, and energy sources. This observation is the driving insight that
defines the scope of this dissertation. By synergistically integrating renewable energy
characteristics with emerging hardware/software supports at different layers of the
hierarchical structure of computing infrastructures, one can greatly improve the
demand-supply coordination effectiveness, and ultimately open the door for a new class
of sustainable high-performance computing platform.
13
CHAPTER 1 INTRODUCTION AND BACKGROUND
The rapid adoption of cloud computing and data analysis is deemed a powerful
engine for the growth of installed server capacity. Data centers housing thousands of
servers have become essential to the operation of businesses, academic, and
governmental institutions. To support emerging data-analytic workloads that tend to
scale well with large number of compute nodes, modern datacenters are continually
adding computing resources (i.e., scaling out) to their existing sites. In turn, the global
server market is projected to triple in 2020, accounting for over 1000 terawatt hour (109
kWh) annual energy consumption [1] [2].
Consequently, the environmental impact of information technology (IT) has
become a growing concern worldwide. In the past ten years, Google server’s electricity
demand has increased almost 20-fold [2]. The worldwide data centers also run the risk
of doubling their energy consumption every 5 years [2]. These large-scale computing
infrastructures, if operated entirely using power generated from conventional fossil fuel,
will unavoidably increase the total cost of ownership (TCO) and irreparably damage the
environment. According to a McKinsey Quarterly report [3], the annual CO2 emissions
of computing systems will research 1.54×109 metric tons within eight years, which could
make IT company among the biggest greenhouse gas emitters by 2020. Consequently,
many environmental activist groups such as the Greenpeace [4] have called on data
center operators to make renewable energy a priority.
Faced with a growing concern about the projected rise in both server energy
expenditures and greenhouse gas emissions, academia and industry today are focusing
more attention than ever on the sustainability of data centers. Nevertheless, simply
14
capping peak power demand and total energy consumption in data centers can hardly
improve overall sustainability. On the contrary, it can significantly limit the workload
performance that we will be able to achieve on current and future cloud computing
platforms. This is because the major source of unsustainability – a heavy dependence
on conventional fossil fuel based electricity – still exists in current designs. As a result,
renewable energy (e.g., solar and wind energy) has attracted considerable attentions
from different sectors in the IT industry. Several companies, including Microsoft, IBM,
Google, Facebook, Apple, HP, and eBay, all start to explore renewable energy
integration in data centers as part of their long-term energy strategies and corporate
environmental goals. According to their plans, these enterprise data centers will draw
power from onsite renewable energy generators, nearby co-located clean energy plants,
or a combination of both.
While it is fairly easy to incorporate green energy sources into a data center’s
energy portfolio, managing renewable energy powered computing system can be a
great undertaking. The variability of renewable power generation, together with the
stochastic workload fluctuation, can cause varying degrees of power mismatch between
the supply and demand. Due to their sensitivity to power disturbances, computing
systems must maintain a continuous balance between the fluctuating renewable power
budget and the variable IT power demand. The power mismatch can have a significant
adverse impact on energy efficiency and productivity. Conventionally, one can smooth
out the temporal variation in renewable power with substantial support from the utility
grid or uninterruptable power supplies (UPS). However, increasing the standby capacity
of utility grid and over-provisioning the UPS system often increase the cost overhead
15
significantly. In addition, the power mismatch problem can cause frequent and
excessive battery discharging events, which not only increase UPS failure rate but can
also quickly deplete the stored energy that is crucial for handling emergencies. Further,
both grid-inverter and UPS battery incur significant efficiency problems. For example,
the round-trip energy loss of battery is typically around 25% [5]. Therefore, appropriate
power management and workload scheduling is the crux of eliminating power
disruptions and ensuring high quality of services.
This dissertation focuses on the question of how to leverage clean energy
technologies and emerging hardware and software techniques to tackle the grand
energy and environmental challenges faced by the power-hungry computing systems
today and tomorrow. The driving insight of this work is the observation that the main
obstacle of green energy integration in computing systems is the lack of cross-
layer power management solutions that spans workload application, computer
hardware, and energy sources.
In the past, energy-efficient computing system designs only emphasize reducing
load power consumption and maximizing resource utilization. They are not aware of the
intrinsic behaviors of renewable power supply, and consequently overlook the interplay
between power systems and computing systems. By synergistically integrating
renewable energy characteristics with novel hardware/software supports, one can
greatly improve the demand-supply coordination effectiveness, and ultimately open the
door for a new class of sustainable high-performance computing platforms.
16
CHAPTER 2 RELATED WORK
While energy-aware computer system has become a research focus since long
time ago, designing green energy (e.g., solar energy and wind energy) powered
computing systems gained its popularity only recently.
Conventional power management techniques for processors and server systems
all assume constant power supply [6]. The goal of these techniques is either reducing
load power consumption or maximizing the performance under fixed power budget. For
example, Teodorescu et al. [7] propose a power management scheme that uses linear
programming to optimize the average throughput with given power budget. Isci et al. [8]
evaluates different power management policies for a given power budget. Lee et al. [9],
analyzes the throughput impact of applying dynamic voltage and frequency scaling
(DVFS) and per-core power gating to power-/thermal- constrained multi-core processors.
Different from conventional energy source, renewable power generators are not a
constant power source. Existing proposals that aim at improving energy efficiency under
fixed power budget cannot guarantee the same energy utilization when the power
budget is variable and unpredictable.
At the data center level, three different renewable energy management
approaches have emerged in recent years. The first design approach focuses on
hardware control techniques. For example, Blink [10] dynamically adjusts the on/off
power cycles of server motherboard to match server power demand to wind power
budget. Deng et al. [11] leverages distributed grid-tie inverters for managing renewable
power integration in data centers. In contrast, the second approach leverages workload
adaptation schemes [12] [13]. The main idea is to shift deferrable workloads to a time
17
window in which renewable generation is sufficient (temporal adaptation), or to relocate
workloads to a different system where power budget is abundant (spatial adaptation).
For the third approach, the gap between power supply management and workload
management starts to diminish. By jointly tune both energy sources and workloads, one
can achieve a better overall performance and efficiency [14].
Many recent papers have discussed the role of battery in server clusters and data
centers. For example, Govindan et al. [15] investigates the use of energy storage
(particularly the UPS batteries) to manage data center peak power. In [16], the authors
also use UPS batteries as the major tuning knob for minimizing datacenter power cost
in aggressively under-provisioned datacenter. Kontorinis et al. [17], has explored the
benefits of deploying distributed UPS system architecture in data centers.
Several recent proposals focus on optimizing cost and energy utilization in green
data centers. For example, Liu et.al [18] models and evaluates dynamic load scheduling
scheme in geographically distributed systems for reducing data center electricity prices.
Deng et.al [19] explores algorisms for optimizing clean energy cost for green Web
hosts. Ren et.al [20] demonstrates that intelligently leveraging renewable energy can
lower data center costs, not just cap carbon footprints.
In addition, several studies have demonstrated the feasibility of renewable energy
powered data centers. These designs typically employ energy storage devices, grid-tie
power controller, or a combination of both to manage renewable power. For example,
HP Labs [21] tests a renewable energy powered micro-cluster called Net-Zero for
minimizing the dependence on traditional utility grid. Their scheduling considers
shedding non-critical workload to match the time-varying solar energy output. In [14],
18
the authors have developed a workload and energy source co-management scheme on
a prototype green data center called Parasol. This work highlights data center free
cooling, low-power server nodes, and net-metering mechanism.
This dissertation explores efficient renewable energy integration at different layers
of the hierarchical structure of modern data centers. In contrast to prior work, it looks at
four novel system architecture with fundamentally different design perspectives: 1) a
solar energy powered multi-core processor that aims at jointly improving renewable
energy utilization and workload throughput [22]; 2) a wind energy powered server
cluster that aims at harvesting wind energy utilization while minimizing power
management overhead [23]; 3) a distributed generator powered data center that targets
near-optimal workload performance without sacrificing system reliability [24]; 4) a
modular green data center design that enables modern power-/carbon- constrained
data centers to scale out economically and sustainably [25].
19
CHAPTER 3 SERVER-LEVEL RENEWABLE POWER MANAGEMENT:
RETHINKING LOAD ADAPTATION SCHEMES
At the server system level, a cross-layer power management scheme that can
jointly manages power supply and server load is of paramount importance.
Although solar energy is an abundant energy source, its efficient utilization
presents a new challenge. A photovoltaic (PV) cell (a.k.a. solar cell is the basic building
block of PV systems, where the electromagnetic radiation of sun can be directly
converted to electricity through the photovoltaic effect [26]. PV cells have a non-linear
relationship between their output photocurrent and terminal voltage. Under uniform
irradiance (solar energy per unit area of the solar panel’s surface) and temperature,
photovoltaic arrays exhibit a current-voltage characteristic with a unique point, called the
maximum power point (MPP) [26], where the module produces maximum output power.
Due to the high fabrication cost and low conversion efficiency (typically 13~19%) of
present solar cells, it is crucial to operate the solar panel at the MPP since doing so can
help PV system users recover their investment more quickly.
Nevertheless, the power generated by PV systems changes over time due to the
unpredictable nature of weather pattern. The unstable working environment along with
the existence of optimum operating points necessitate the rethinking of server power
management policy to achieve higher efficiency in utilizing solar energy. Specifically,
this work focuses on multi-core servers which are the mainstream hardware design
choice for today’s IT industries and demand increasing amount of green power to
unleash their full computation potential. The multi-core processor load should be
regulated and tuned in such a way that the overall workload performance is also
optimized during the solar energy harvesting process.
20
Figure 3-1. Three typical PV systems. A) Grid-connected systems. B) Direct-coupled systems. C) Battery-equipped systems.
There are various ways to connect solar panels to a computer load, as shown in
Figure 3-1. For instance, solar-powered systems can employ energy storage elements
(e.g., batteries or super-capacitors) [27] to reduce voltage fluctuations and perform
maximum power tracking. However, energy storage elements introduce many
limitations in PV system. For example, the maximal energy that can be delivered is
limited by the battery capacity and this unavoidably affects the performance of multi-
core systems. The large current drawn by multi-core chips requires battery with large
capacity which is both bulky and expensive. In addition, the turn-around efficiency of
battery is about 75% due to internal resistance and self-discharge [5]. It causes
significant loss in overall solar energy utilization. Last but not the least, existing
rechargeable batteries all have limited lifetime. Frequent charge/discharge and self-
discharge further aggregate the aging effect. Without proper maintenance (which
requires human intervention and causes loss of availability), the aging problem will
directly lead to the capacity reduction and output voltage change. As a result, over the
lifetime of the solar-powered systems, batteries cost (e.g. initial and maintenance cost)
can be the most expensive component of a renewable energy system [28].
In this study we consider cost-effective direct-coupled PV systems. Different from
the conventional design, we propose techniques that allow multi-core processors to
autonomously harvest renewable energy to the maximal degree.
21
A B
Figure 3-2. The current-voltage (I-V) and power-voltage (P-V) characteristics of PV cells. A) I-V curve and MPP. B) P-V curve and MPP.
3.1 PV Systems and Load Matching
For a given current-voltage curve (i.e., I-V curve), there is a single operating point
where the values of the voltage and current of the cell result in the maximum power
output, which is the optimal operation point (maximum power point, MPP) for the
efficient use of the solar energy. As shown in Figure 3-2, a load connected PV cell may
operate (as determined by the intersection of the I-V curve and the load line) far from its
optimal condition, resulting in inefficient solar power utilization.
When connected to electrical loads, solar panels generally employ power
converters to achieve appropriate output voltage. Assuming 𝑃𝑖𝑛 = 𝑃𝑜𝑢𝑡, the converters
can be described as 𝑉𝑜𝑢𝑡 = 𝑉𝑖𝑛/𝑘 and 𝐼𝑜𝑢𝑡 = 𝑘 ∙ 𝐼𝑖𝑛, where k is the transfer ratio. The
actual operating point of the PV system occurs at the intersection of the electrical
characteristics of the solar panel and the load. By load matching, a maximum power
point tracker (MPPT) is able to extract the optimal power from the panel under varying
atmospheric conditions [26]. Such load matching can be achieved by either adjusting
the DC-DC converter transfer ratio, k, or tuning multi-core load, w. For a multi-core
processor, its electrical characteristics are largely dependent on parameters such as
clock frequency and supply voltage. For instance, as clock frequency increases, the
processor will exhibit lower impedance and draw more current and power.
22
A B
Figure 3-3. The characteristics of load matching. A) I-V curve with operating point on the right. B) I-V curve with operating point on the left.
Figure 3-3 shows how the operating point changes if we tune the multi-core load
w and transfer ratio k. The PV output voltage decreases when the load line moves
counterclockwise and increases when the load line moves clockwise. However, the
actual output power may either increase or decrease as the load line moves, depending
on whether the operating point is approaching the MPP or not.
Note that simply increasing the multi-core processor utilization rate and voltage
level will not automatically draw maximal power because the required load adaptation
scheme (i.e. increasing or decreasing the load) varies at different operating points. On
the other hand, tuning the transfer ratio alone cannot increase the multi-core
performance either due to the lack of effective multi-core load adaptation schemes.
Moreover, without appropriate coordination between the power converter and load
adaptation, the system will suffer from severe voltage fluctuations.
3.2 Managing Solar Energy Powered Multi-core System
SolarCore is a multi-core architecture power management scheme that can
dynamically track the maximal solar energy generation while improving workload
performance. One of the key benefits of using SolarCore is that it reduces the power
dependence on conventional fossil-fuel-based power supply.
23
Figure 3-4. An overview of SolarCore power management architecture.
Figure 3-4 provides an overview of the SolarCore architecture. The system is
powered by solar energy with utility grid as backup. An automatic transfer switch (ATS)
is employed to seamlessly select between the primary (i.e. the solar panel) and backup
power sources and an uninterruptible power supply (UPS) ensures continuous power
delivery to the load. An AC/DC converter is used only if the ATS switches to the utility. A
tunable power-conservative matching network is used to convert the PV output voltage
to the level that is compatible to existing multi-core processors and systems.
SolarCore assumes that the processor, which contributes to the significant power
dissipation in typical multi-core systems, is the only component powered by the
renewable energy while the rest of the system is power by utility grid. At the front end,
both load current and voltage are measured via I/V sensors and the results are fed-back
to a SolarCore controller, which is responsible for identifying and tracking the maximal
power point. The controller adjusts the DC/DC converter through MPPT control signal
and communicates with the processor through an adapter, which is responsible for per-
core load tuning and workload optimization. The harvested solar energy is allocated
across all running cores with the objective of maximizing the overall performance.
24
SolarCore decomposes solar energy management into 1) multi-core aware MPP
tracking and 2) dynamic server load adaptation. It uses per-core dynamic voltage and
frequency scaling (DVFS) [29] and per-core power gating (PCPG) techniques for load
adaptation. The purpose of this load adaptation is to move the operating point of solar
array closer to the MPP under changing atmospheric conditions. To achieve per-core
DVFS, SolarCore leverages on-chip voltage-regulator modules (VRM) [30] to provide
the supply voltage for each core. The supply voltage information is communicated
among the SolarCore controller, the VRMs, and cores via a number of bits called the
Voltage Identification Digital (VID), which has been implemented in commercial
microprocessors. For example, the Intel Xeon processor employs a 6-bit VID to specify
the input voltage from 0.84V DC to 1.6V DC with 32 different voltage steps.
3.2.1 Multi-core Aware MPP Tracking
The multi-core aware MPP tracking of SolarCore relies on successive tuning of
both the DC/DC converter transfer ratio k and the multi-core load w. The SolarCore
controller aims at coordinating the power supply converter and the multi-core load
adaptation so as to achieve maximal power drawn. In addition, this technique ensures
1) a correct tracking direction and 2) a stable load operating voltage 𝑉𝑙.
Table 3-1 summarizes the electrical behavior of load tuning. Tuning k or w can
increase output power 𝑃𝑜𝑢𝑡, load voltage 𝑉𝑙 and load current 𝐼𝑙, depending on the
location of the operating point. By simultaneously adjusting k and w, one can increase
the output current while maintaining constant output voltage. As a result, the operating
point moves closer to MPP and the processor can utilize more power. SolarCore uses a
three-step control to perform MPP tracking, as shown in Figure 3-5.
25
Table 3-1. The impact of load tuning on output power behaviors. The position of current operating point affects the way we perform MPP tracking.
Position MPPT Operation Power Load voltage Load current
Left of MPP 𝑘 ↑ (𝑐𝑜𝑛𝑠𝑡. 𝑙𝑜𝑎𝑑) 𝑃𝑜𝑢𝑡 ↑ 𝑉𝑙 ↑ 𝐼𝑙 ↑
Left of MPP 𝑤 ↑ (𝑐𝑜𝑛𝑠𝑡. 𝑡𝑟𝑎𝑛𝑠𝑓𝑒𝑟 𝑟𝑎𝑡𝑖𝑜) 𝑃𝑜𝑢𝑡 ↓ 𝑉𝑙 ↓ 𝐼𝑙 ↑
Left of MPP 𝑘 ↑ + 𝑤 ↑ 𝑃𝑜𝑢𝑡 ↑ ∆𝑉𝑙 ≈ 0 𝐼𝑙 ↑
Right of MPP 𝑘 ↓ (𝑐𝑜𝑛𝑠𝑡. 𝑙𝑜𝑎𝑑) 𝑃𝑜𝑢𝑡 ↑ 𝑉𝑙 ↑ 𝐼𝑙 ↑
Right of MPP 𝑤 ↑ (𝑐𝑜𝑛𝑠𝑡. 𝑡𝑟𝑎𝑛𝑠𝑓𝑒𝑟 𝑟𝑎𝑡𝑖𝑜) 𝑃𝑜𝑢𝑡 ↑ 𝑉𝑙 ↓ 𝐼𝑙 ↑
Right of MPP 𝑘 ↓ + 𝑤 ↑ 𝑃𝑜𝑢𝑡 ↑ ∆𝑉𝑙 ≈ 0 𝐼𝑙 ↑
Figure 3-5. The flowchart of SolarCore maximum power point tracking.
Step 1. The algorithm starts with a normal operating voltage (i.e. 𝑉𝑙 = 𝑉𝑑𝑑). Due
to supply variation, 𝑉𝑙 may not equal to 𝑉𝑑𝑑 at each beginning phase of periodically
triggered MPP tracking. In this step, the SolarCore controller will restore the output
voltage to 𝑉𝑑𝑑 by decreasing or increasing the load to an appropriate level. This step
avoids system over-loading and serves as preparation for other steps.
Step 2. As Table 3-1 shows, the transfer ratio tuning direction that approaches
MPP depends on the system’s current operating position. To determine the tuning
direction, our techniques set the transfer ratio of the DC/DC converter from 𝑘 + ∆𝑘 to
and observe the current. An increase in output current suggests that the panel
generates more power and the system is approaching to the MPP. In this case, the
actual operating point is on the left side of the MPP and our algorithm proceeds to step
(3) for load matching. On the other hand, a decrease in output current suggests a wrong
26
tuning direction. In this case, the system further decreases the transfer ratio by 2∆𝑘
which results in a net change of −∆𝑘 in transfer ratio. Consequently, our technique
resumes the correct direction and proceed to step (3) to perform load matching.
Step 3. In this step, we tune the load until 𝑉𝑙 equals 𝑉𝑑𝑑. Due to the adjustment of
the transfer ratio in step (2), the output voltage changes as well. By increasing the load,
we can change the PV operating point and decrease 𝑉𝑙 until it reaches 𝑉𝑑𝑑. During each
control period, SolarCore increases the load successively with the aid of tuning transfer
ratio k as discussed in steps (2) and (3). Through the stepwise tuning of transfer ratio
and load matching, this method progressively adapts multi-core power consumption and
eventually reaches to the new MPP. The goal of SolarCore power management is
appropriate coordination of the variable power supply and demand.
3.2.2 Dynamic Per-Core Load Tuning
SolarCore uses DVFS to perform successive load tuning. To be more specific, it
applies per-core DVFS to manage multi-core power at a fine granularity and achieve
more effective maximum power tracking. There has been prior work [7] that formulates
DVFS control as linear programming problem and using online linear programming to
optimize power and performance trade-offs has been proven to be efficient.
The processor dynamic power model used in SolarCore is 𝑃 = 𝛼𝐶𝑉2𝑓 [6]. Our
system further assumes that: (1) the voltage scales approximately linearly with
frequency within typical operating ranges, i.e., 𝑓𝑖 = 𝜇𝑉𝑖 + 𝜆, 𝑉𝑖 ∈ {𝑉1, 𝑉2, … , 𝑉𝑛}, where n is
the number of voltage levels; (2) the total power drawn by a core is approximately a
cubic function of its voltage level, i.e. 𝑃𝑖 = 𝛼𝐶𝑉𝑖2(𝜇𝑉𝑖 + 𝜆) ≈ 𝑎𝑖𝑉𝑖
3+ ci; (3) the voltage
scaling has little impact on IPC and the throughput 𝑇𝑖 = 𝐼𝑃𝐶𝑖 × 𝑓𝑖 of a core is
27
proportional to its frequency; and (4) the available power for an individual core does not
exceed its maximum power and all cores are homogeneous. Given the above
assumptions, the per-core throughput can be expressed as 𝑇𝑖 = 𝐼𝑃𝐶𝑖 × (𝜇𝑉𝑖 + 𝜆) =
𝑏𝑖𝑉𝑖 + 𝑑i, and the average throughput across all cores is 𝑇𝑖̅̅ ̅ =
1
𝑁∑(𝑏𝑖𝑉𝑖 + 𝑑i), where N is
total number of cores. Since the system working duration, which is subject to the solar
power generation level, plays a key role in determining the overall performance, our
goal is to maximize 𝑇𝑖̅̅ ̅ × 𝑅𝑢𝑛𝑡𝑖𝑚𝑒/𝑑𝑎𝑦. This performance-time product (PTP) can be
measured as the total instructions committed per day. Intuitively, to maximize the PTP,
one needs to increase both the throughput and the system operation duration.
In contrast to conventional power management problem, solar power is not fixed,
i.e., 𝑃𝑡𝑜𝑡𝑎𝑙 = ∑ 𝑃𝑖 = ∑(𝑎𝑖
𝑉𝑖3 + 𝑐𝑖) ≤ 𝑃𝑀𝑃𝑃 and 𝑃𝑀𝑃𝑃 = 𝑓(𝐼𝑟𝑟𝑎𝑑𝑖𝑎𝑡𝑖𝑜𝑛, 𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒). The
unpredictable nature of PV output makes performance optimization much more
challenging. In this work, SolarCore uses a throughput-power ratio (TPR) optimization –
a PTP optimization method that can be performed along with load adaptation. The
throughput-power ratio specifies the throughput speedup of a processor when additional
power is available. When adjusting the core voltage by ∆𝑉, a power consumption
change of ∆𝑃 = 3𝑎𝑖𝑉𝑖2∆V will occur. The relationship between throughput change and
voltage scaling is given by ∆𝑇 = 𝑏𝑖∆𝑉. Hence the throughput power ratio can be defined
as ∆𝑇
∆𝑃=
𝑏𝑖
3𝑎𝑖𝑉𝑖2 . Both 𝑏𝑖 and 𝑎𝑖 can be derived using profiling information obtained from
the performance counters and the current/voltage sensors at each beginning of MPP
tracking. Given the throughput power ratio, our scheme identifies the appropriate core
voltage setting through a heuristic process. The cores that exhibit large throughput-
power ratio will have higher priority to receive the available power.
28
SolarCore stores the core ID and per-core voltage level in a table sorted by the
throughput-power ratio, as shown in Figure 3-6. As described previously, load matching
is achieved by increasing or decreasing the voltage level and frequency of the selected
core. When we increase the load, we choose the core with higher throughput-power
ratio, which will maximize the performance by using additional power. When the load
needs to be decreased, our scheme chooses the core with lower throughput-power
ratio. Doing so allows us to meet the power budget while minimizing the impact on the
overall system performance. The above step is performed iteratively until the
aggregated multi-core power approximates the new power budget.
Figure 3-7 shows the per-core load adaptation scenario. The optimum computing
load is achieved when the inflection point (the point where the slop of power trace
changes sign) is reached. Since the inflection point can be stable operation states or
transition states (i.e. depending on whether the inflection point is under the power
budgets), the MPPT controller should not stop tuning immediately once an inflection
point is met. For example, if the inflection point is above the power budget, the MPPT
controller needs to decrease the load further. As a result, the system returns back to the
stable load level and the optimum power consumption is met.
Note that there is still power headroom even if we perform maximum power
budget racking. However, the existence of such a power margin is actually beneficial
since it improves the robustness of the system. Figure 3-8 shows the pseudo code of
our per-core load-tuning algorithm. The processor starts tuning its load when the
controller detects a change in PV power supply. Successive server load adaptation is
performed at the beginning of each triggered tracking period.
29
Figure 3-6. The per-core load adaptation policies of SolarCore.
Figure 3-7. The per-core load adaptation demonstration of SolarCore.
Figure 3-8. Pseudo code for SolarCore per-core load adaptation.
3.3 Experimental Methodology
SolarCore is evaluated through heavily extended simulators and real-world solar
traces. Our PV power model takes irradiance and temperature profiles as input and
generates the I-V and power profiles. Table 3-2 shows the meteorological data from the
Measurement and Instrumentation Data Center (MIDC) [31] of the National Renewable
Energy Laboratory. The MIDC provides real-time records from measurement stations
located throughout the US. Table 3-2 lists different locations that have different solar
energy resource potentials. For each location, we evaluated our design with daytime
solar irradiance data (i.e. 7:30AM ~ 5:30 PM) across different seasons (e.g. the middle
of Jan., Apr., Jul. and Oct. in the year of 2009). We trigger the maximum power point
tracking at each of 10-minute interval. We observed that the tracking duration within
each interval is less than 5ms. We applied sampling techniques to capture the
irradiance changes at a large time scale.
30
Table 3-2. The evaluated different solar power traces. Station
Location
KWh/m2/day on Average Solar Energy Potential
PFCI Phoenix, AZ > 6.0 Excellent
BMS Golden, CO 5.0 ~ 6.0 Good
ECSU Elizabeth City, NC 4.0 ~ 5.0 Moderate
ORNL Oak Ridge, TN < 4.0 Low
Table 3-3. Performance levels of battery-equipped PV systems.
Level Range MPP Efficiency Battery Efficiency Overall de-rating factors
High 81~ 92% 97% 95% 92% Moderate 70~ 81% 95% 85% 81% Low < 70% 93% 75% 70%
We also compared the total solar energy drawn by SolarCore with that of battery-
equipped standalone PV system. We assume that the battery is optimally charged using
MPPT circuit and the multi-core processor runs with full speed using stable power
supply. We assume that dynamic power monitor is used to ensure that all the solar
energy stored in the battery is consumed in our simulation. We use de-rating factors to
estimate the upper bound of the utilization on conventional battery-equipped solar
power systems. Typically, battery loss ranges from 5% to 25% [5] and a modern
maximum power tracking controller has conversion efficiency between 93% ~ 97%. We
consider three levels of performance ranges, as shown in Table 3-3.
We simulated a multi-core system comprised of 8 Alpha 21264 processors where
each core has private L1 and L2 caches. We assume a frequency of 2.5GHz and a
90nm technology with a maximum supply voltage of 1.45V. Table 3-4 summarizes our
core configuration. We used a cycle-accurate multi-core simulator integrated with the
modified power models from Wattch [32] and CACTI [33]. Both dynamic and leakage
power are modeled. We assume that the dynamic voltage and frequency scaling
schemes employed in each core are similar to Intel’s SpeedStep techniques [34]: each
core supports 6 frequency and voltage operating points and the frequency can be
31
scaled from 2.5GHz down to 1GHz with a 300MHz stepping. The voltage scales
approximately linearly with frequency within typical operating ranges; six voltages
ranging from 1.45V to 0.95V with a 0.1V stepping are used. To support DVFS, the
Wattch and CACTI power models are extended to use the voltage and frequency as
input and then adjust the power estimation for a core when the operating status
changes. For the performance analysis, we used the performance-time products rather
than IPC since the frequency varies across the entire program execution.
We use workloads from the SPEC2000 benchmark suite shown in Table 3-5. To
form multi-programming workloads, we use a program’s average energy-per-instruction
(EPI) and categorize the benchmarks as being high-EPI (EPI 15nJ), moderate-EPI
(15nJ EPI 8nJ), or low-EPI (EPI 8nJ). We run each benchmark in their
representative execution intervals and the EPI is obtained by calculating the average-
energy consumed per-instruction. We generated both homogeneous and
heterogeneous workloads, as shown in Table 3-5.
Table 3-4. Simulated SolarCore machine configuration. Parameters Configurations
Frequency 2.5/2.2/1.9/1.6/1.3/1.0GHz Voltage 1.45/1.35/1.25/1.15/1.05/0.95V Width 4-wide fetch/issue/commit Issue Queue 64 ITLB 128 entries, 4-way, 200 cycle miss Branch Predictor 2K entries Gshare10-bit global history BTB 2K entries, 4-way Return Address Stack 32 entries RAS ROB Size 98 entries LSQ 48 entries Integer ALU 4 I-ALU, 2 I-MUL/DIV, 2 Load/Store FP ALU 2 FP-ALU, 2 FP-MUL/DIV/SQRT DTLB 256 entries, 4-way, 200 cycle miss L1 I-/D- Cache 64KB, 4-way, 64 B/line, 2 ports, 3 cycle access L2 Cache 2MB, 8-way, 128 B/line, 12 cycle access Memory Access 64 bit wide, 400 cycles access latency
32
Table 3-5. The simulated multi-programmed workloads. EPI Value Workload Set 1 Workload Set 2
High H1(art×8) H2 (art×2, apsi×2, bzip×2, gzip×2) Moderate M1(gcc×8 ) M2 (gcc×2, mcf×2, gap×2, vpr×2) Low L1(mesa×8) L2 (mesa×2, equake×2, lucas×2, swim×2) High-Moderate HM1(bzip×4, gcc×4) HM2 (bzip, gzip, art, apsi, gcc, mcf, gap, vpr) Moderate-Low ML1(gcc×4, mesa×4) ML2 (gcc, mcf, gap, vpr, mesa, equake, lucas, swim)
Table 3-6. The evaluated solar power management schemes. Algorithm MPPT Load Tuning Scheduling Method
Fixed-Power No DVFS Linear programming with a fixed power budget MPPT & IC Yes DVFS Concentrating solar power onto a few cores MPPT & RR Yes DVFS Round-robin scheduling MPPT& Opt Yes DVFS Optimized scheduling based on throughput-power ratio
3.4 Evaluation Results
We compare the efficiency of SolarCore with various power management
policies, as summarized in Table 3-6. The Fixed-Power is a non-tracking power
management scheme which assumes a constant power budget during the entire
workload execution. In contrast to SolarCore, MPPT&IC and MPPT&RR both apply
tracking control techniques but employ different scheduling policy to perform load
adaptation. MPPT&IC keeps tuning individual core until reaching its highest or lowest
V/F level, while MPPT&RR distributes the power budget variation evenly across all the
cores in a round-robin fashion. The MPPT&Opt (i.e. default configuration for SolarCore)
selects cores using the proposed throughput-power ratio optimization method. We also
compared the power efficiency to that of the battery-equipped PV systems.
The operation duration we refer to in this work will always be the daytime
duration. In addition, the effective operation duration means the duration that multi-core
processor successfully performs MPP tracking and workload optimization with
SolarCore power management technique. Due to the intermittency of solar energy,
effective operation duration will always be shorter than the operation duration.
33
A B C
Figure 3-9. MPP tracking results for different workloads in a sunny day. A) Tracking result for H1. B) Tracking results for HM2. C) Tracking results for L1.
A B C
Figure 3-10. MPP tracking results for different workloads in a cloudy day. A) Tracking result for H1. B) Tracking results for HM2. C) Tracking results for L1.
Figure 3-11. The histogram of the power tracking error.
3.4.1 Power Tracking Accuracy
Figures 3-9 and 3-10 provide a graphic view of the tracking results. Each figure
shows the maximum power trace and the actual power harvested with tracking. We
present results for sunny and cloudy weather patterns. For high EPI workloads (e.g.
H1), large ripples in power tracking are generated due to the high variation in load
power. Power ripples, together with the unpredictable nature of the environment, can
affect the use of renewable energy and the system reliability. Low EPI workloads and
heterogeneous workloads are observed to have small ripples. The former manifests
small power ripples since they have lower average energy per instruction, which results
in relatively small power fluctuation amplitude; the latter exhibits small ripples because a
mix of diverse workloads can smooth the power demand fluctuation.
100 200 300 400 500 6000
20
40
60
80
100
Time (min)
Pow
er
(watt)
Maximal Power Budgets
Actual Power Consumptions
100 200 300 400 500 6000
20
40
60
80
100
Time (min)
Pow
er
(watt)
Maximal Power Budgets
Actual Power Consumptions
100 200 300 400 500 6000
20
40
60
80
100
Time (min)
Pow
er
(watt)
Maximal Power Budgets
Actual Power Consumption
100 200 300 400 500 6000
50
100
150
Time (min)
Pow
er
(watt)
Maximal Power Budgets
Actual Power Consumption
100 200 300 400 500 6000
50
100
150
Time (min)
Pow
er
(watt)
Maximal Power Budgets
Actual Power Consumptions
100 200 300 400 500 6000
50
100
150
Time (ms)
Pow
er
(watt)
Maximal Power Budgets
Actual Power Consumption
34
A B C
Figure 3-12. Operation durations on different power budget thresholds. A) Declines slowly. B) Declines linearly. C) Declines rapidly.
Our dynamic load adaptation scheme reduces the impact of load power ripples
by leaving a small power margin between the actual load power consumption and the
solar power budget. This power margin slightly degrades the tracking accuracy but
improves the reliability. To quantify the impact of power margin on the tracking
accuracy, we calculated the relative tracking error. In each tracking period t, the relative
tracking error is defined as �̅� = |𝑃 − 𝑃′| 𝑃′⁄ , where 𝑃 is the actual power consumption
and 𝑃′ is the power budget. Figure 3-11 shows the histogram of power tracking errors
across 12 locations (different weather patterns) and 10 workload sets. Due to the
reduced power margin, the low EPI workloads (e.g. L1) show relatively small tracking
errors [22]. Similarly, compared with H1, heterogeneous workloads (e.g. HM2) have
small tracking error as well [22]. The majority of tracking errors fall into 8% to 16%.
3.4.2 Fixed Power Budget
For the direct-coupled PV systems, the load starts to operate using solar power
when the amount of renewable energy exceeds a power-transfer threshold. We
evaluated the solar energy utilization on multi-core systems that use power-transfer
threshold as a fixed power budget. When the renewable power output falls below the
power-transfer threshold, we switch to utility grid.
35
A A
B B
C C
D D
Figure 3-13. Normalized solar energy usage under different fixed power budget. A) Arizona. B) Colorado. C) North Carolina. D) Tennessee. (x-axis: watts)
Figure 3-14. Normalized performance time product (PTP) under different fixed power budget. A) Arizona. B) Colorado. C) North Carolina. D) Tennessee. (x-axis: watts)
We simulated both homogeneous and heterogeneous workloads with high and
low EPI. For each weather pattern, we calculated the effective operation duration under
different power budget thresholds, as shown in Figure 3-12. The threshold (x-axis:
watts) affects processor performance-time product (PTP) in both throughput and
effective operation duration. A higher power-transfer threshold will make the multi-core
processor run at higher voltages and frequencies but only for short durations. A
36
conservative, low power-transfer threshold operating mode will have longer duration but
lower clock frequencies. Intuitively, the processor should run under higher (lower) power
budget to maximize its performance when the effective operation duration declines
slowly (rapidly). Extensive simulations shown in Figures 3-13 and 3-14 further proof this.
We calculated the accumulated energy drawn by the multi-core processor and
the performance in terms of PTP. All the results are normalized to those obtained on
SolarCore. As shown in Figures 3-13 and 3-14 (average results across all benchmarks),
the maximal solar energy drawn does not guarantee maximal performance (e.g. July
@CO and Apr @TN). The maximal workload performance may occur under a high
power budget (e.g. Apr @AZ), a moderate power budget (e.g. Jan @TN) or even a low
power budget (e.g. Apr @NC). Therefore, a single, optimal fixed power budget for the
multi-core system does not exist. Even under the optimal fixed power budget, the best
energy utilization and PTP that the Fixed-Power schemes can achieve is less than 70%
of that yielded on maximum power tracking. In other words, SolarCore outperforms
Fixed-Power control scheme by at least 43% (i.e. (100%-70%)/70%) in terms of both
energy utilization and workload performance.
3.4.3 Solar Energy Utilization
The energy utilization of SolarCore depends on the solar resource potential of
the geographic location. We calculated the energy utilization (i.e. actual solar energy
consumed / overall solar energy generated) with various load adaptation scenarios. As
shown in Figure 3-15, the average energy utilization drops when the renewable energy
potential is low. For locations with abundant solar resource, SolarCore draws 5% more
power compared to a typical battery equipped PV system which has an energy
utilization upper bound of 81%.
37
Figure 3-15. Energy utilization across different locations. The upper bound of energy
utilization on battery-based system is estimated using de-rating factors.
In addition to the weather patterns, the load adaptation methods and workload
characteristics affect the utilization as well. In Figure 3-15, MPPT&Opt load adaptation
method has 2.6% lower energy utilization compared with MPPT&RR. This is because
MPPT&Opt relies on successively tuning low-power, high-throughput cores to improve
performance. Compared with the high EPI workloads, low EPI workloads exhibit higher
energy utilization due to the fine-grained load tuning and the reduced power margin.
3.4.4 Performance Improvement
We compared SolarCore with the high efficiency battery-equipped systems.
Battery-L denotes the lower bound of a high-end battery-equipped system which has a
total energy conversion efficiency of 0.81. Battery-U denotes the upper bound of a high-
end battery-equipped system which has a total energy conversion efficiency of 0.92.
Figure 3-10 shows the performance of SolarCore with different load scheduling methods
across different weather patterns. All the results are normalized to Battery-L.
In Figure 3-16, MPPT&IC shows the lowest performance because it concentrates
the solar power into fewer cores. The throughput-power ratio ∆𝑇
∆𝑃=
𝑏𝑖
3𝑎𝑖𝑉𝑖2 shows that the
performance return decreases when we allocate the limited available power to those
high V/F level cores. MPPT&RR increases the performance by distributing the solar
power to each individual core in a round-robin fashion. In this case, each core runs in a
38
moderate V/F level with a relatively high performance per watt. The average
performance of MPPT&IC, MPPT&RR and MPPT&Opt is 0.82, 1.02 and 1.13,
respectively. The normalized performance of Battery-U is 1.14. As can be seen, with
runtime optimization, MPPT&Opt improves the performance by 37.8% compared with
MPPT&IC and 10.8% compared with MPPT&RR. Compared with the most efficient
battery-based systems, SolarCore yields less than 1% performance degradation.
Figure 3-16. Normalized PTP with different load scheduling methods.
39
3.5 Summary of SolarCore Design
Photovoltaic power generation are one of the most promising technologies for the
future of green computing in IT infrastructures. SolarCore represents the first step on
exploring the utilization of solar energy for powering multi-core processors, which are
the core hardware building blocks of current and future IT infrastructures and demand
increasing amounts of power to unleash their computation potential. Unlike conventional
energy sources, PV modules exhibit non-linear I-V characteristics and their output
power changes with the variation of light intensity and ambient temperature. Therefore,
the best design practice of solar energy driven high-performance computing system
requires a cross-layer power management scheme: 1) appropriate power provisioning
control to harvest the renewable energy and reduce the impact of supply variation and
2) dedicated load adaptation schemes that optimize the workload performance.
While existing techniques seek to minimize multi-core power dissipation under
performance constraints or to maximize throughput for a given power budget, SolarCore
harvests the maximum amount of solar energy to maximize the power budget for
optimal throughput without using the expensive storage elements. Furthermore,
SolarCore applies load optimization based on the workload throughput-power ratio to
improve the power application effectiveness. Because of its ability to extract additional
solar energy and its ability for load optimization, SolarCore boosts multi-core processor
performance by 43% compared with conventional fixed power budget scheme and
10.8% compared with relatively straightforward round-robin load adaptation.
40
CHAPTER 4 CLUSTER LEVEL RENEWABLE POWER MANAGEMENT:
UNDERSTANDING THE CONTROL OVERHEAD
This chapter investigates the efficiency issue of incorporating renewable energy
resources into existing utility grid powered server clusters. At the server cluster level, we
primarily focus on wind energy since it is widely used to provide affordable and
abundant green energy for large scale facilities [35].
Due to the intermittent nature of renewable energy, existing designs typically use
on-site renewable generation to compensate part of their total server power
requirement. Consequently, many server clusters can be powered by hybrid energy
sources, i.e., part of the server clusters is powered by wind energy and the remaining
part of the server clusters is powered by utility grid. When renewable energy contributes
a large portion of the load power demand (e.g., > 15%), variations in renewable power
supply can still have a significant impact on load operation. To maximize the utilization
of green energy sources, it is desirable to place as much workload as possible on wind
energy powered clusters. Therefore, dynamic load balancing between the two types of
clusters is important to maintain the desired green energy utilization. However,
aggressively matching server power demand to renewable power budget can degrade
system response time (due to load power capping or control overhead) and provides
very limited energy efficiency return.
Many research efforts are focused on reducing idle server power [36], lowering
provisioned power capacity [37], managing computing power and cooling power [38],
and optimizing power allocation [39] [40]. However, the impact of renewable power
variation behavior on server cluster operation is still an unexplored area.
41
A B
Figure 4-1. Power management overhead vs. renewable energy utilization. A) Normalized control overhead. B) Average energy utilization.
We simulate a wind energy powered server cluster which tracks the renewable
power output surge with a pre-defined coverage factor (CF). It tracks the renewable
power budget whenever renewable generation decreases to avoid brownout. When
CF=1, it means we aggressively track the wind power variability. As shown in Figure 4-
1A, compared to no-tracking (i.e., CF=0), always tracking the power variation (i.e., CF =
1) increases the load tuning overhead by 2X. Nevertheless, the improvement of energy
utilization is less than 15%, as shown in Figure 4-1B.
To improve load matching efficiency, this chapter introduces iSwitch, a light-
weight power management scheme that maintains a desirable balance between
renewable energy utilization and server cluster control overhead. Its novelty is two-fold.
First, it is a supply-aware power tracking mechanism that applies different power
management strategy under different wind variation scenarios. Second, it also features
a load-aware optimization that is able to further minimize the performance overhead.
4.1 Power Management Regions
A wind turbine generates electrical power by extracting kinetic energy from the
air flow. Figure 4-2 shows the output characteristics of a GE wind turbine, whose power
curve is divided into three regions by the designated operating wind speeds. The cut-in
speed is the minimum speed at which the blade starts to rotate. The cut-off speed is the
wind speed at which the turbine shuts down for protecting the blade assembly.
42
Figure 4-2. Wind power output characteristics.
In Figure 4-2, we refer to the three regions as intermittent power outage period
(Region-I), variable power generation period (Region-II) and stable power generation
period (Region-III), respectively. In Region-I, wind power is intermittently unavailable
because the wind speed is either too low or too high. In Region-II, the mechanical
power delivered to the turbine generator is given by 𝑝 = 0.5𝜌𝐴𝑣3𝐶 [35], where 𝜌 is the
air density, A is the swept area of the blades, v is the wind speed and C is called power
coefficient factor. In Region-III, the wind turbine operates at its rated power.
The variations in wind speed are described by the Weibull distribution [35]. In
Equation 4-1, k is the shape parameter, c is the scale parameter and v is the wind
speed. At most wind farm sites, the wind speed has the Weibull distribution with k = 2,
which is specifically known as the Rayleigh distribution [35].
𝑓(𝑣) = (𝑘
𝑐) (
𝑣
𝑐)
𝑘−1𝑒−(
𝑣
𝑐)
𝑘
, 𝑣 ∈ [0, ∞) (Equation 4-1)
As shown in Figure 4-3, the distribution function in Region-II is not monotonic. In
this region, wind speed has equally high possibilities at a wide range of values. As a
result, we are more likely to incur time-varying wind speed in Region-II. In addition, the
turbine output is a steep curve due to the cubic relation between wind power and wind
speed. A small change of the wind speed can lead to large output fluctuation.
43
Figure 4-3. The probability distribution of wind speed.
Figure 4-4. Power variation in wind powered server clusters (x-axis: minutes).
Figure 4-4 shows real traces of wind power generation and server cluster power
demand. It illustrates the aforementioned three scenarios, namely, intermittent wind
power outage period (Region-I), low renewable generation with frequent fluctuation
(Region-II), and full renewable generation with relatively stable output (Region-III).
Similar region partition method is also applicable for intermittent renewable energy
sources such as solar power and tidal power.
Region-I: tune wisely. During the low generation period, it is wise to shift server
cluster load from renewable energy supply side to utility power. Existing practices either
put servers into low power states or apply power cycling techniques [10] on the
hardware. Although these approaches show impressive power control capability, they
sacrifice the computing throughput. In addition, it typically takes a long time (about tens
of minutes as we observed) for renewable energy output to resume. As a result, putting
servers into sleep state and waiting for the renewable energy to resume may not be
wise, especially for parallel clusters that have inter-node workload dependency.
44
Region-II: track wisely. Whenever the load power fluctuates or renewable
energy generation varies, load matching is performed as a common practice to handle
the power discrepancy [10] [22]. In Region-II, the wind generation oscillates severely.
The load power tuning is largely a result of the power supply variation, as shown in
Figure 4-4. However, aggressively matching the load to the supply results in little energy
benefits but disturbs the normal server operation and degrades the performance of
parallel workload. Therefore, seeking appropriate tracking is especially important.
Region-III: schedule wisely. When the renewable energy is stable, frequent
load fluctuation contributes to a number of load tuning operations. In Figure 4-4 (Phase
III), although the data center power has a relatively small dynamic range, frequent
variation invokes a large number of back-and-forth load migration operations (e.g., via
virtual machine migration). Those tuning activities have little contribution to the overall
energy utilization but increase network traffic significantly.
A renewable energy powered data center frequently experiences the
aforementioned three power management regions throughout its lifetime. To improve
overall quality of service, a cooperative power management scheme is required to
handle the above power management regions.
4.2 An Overview of iSwitch Architecture
We propose iSwitch, a server cluster coordination and optimization scheme that
ensures high renewable energy utilization and low operational overhead. It is designed
to provide autonomic load power balancing between conventional utility grid and
renewable energy generation. As shown in Figure 4-5, the on-site renewable power
supply provides the data center with clean energy through a separate power line. We
choose not to synchronize renewable power to the grid because the fluctuating nature
45
of renewable energy often challenges the grid stability [41]. Although one can leverage
dual-corded servers to utilize two power supplies simultaneously, it is not energy-
efficient since the efficiency of power supply units decreases when the computing load
is low [42]. As an alternative approach, iSwitch explores load migration.
The basic idea behind iSwitch is switching, or the operation of performing a
switch. In this work, the switch is defined as a load migration event that leads to
redistribution of load power between two different group of server clusters (and
consequently, between different energy sources) Instead of throttling load power when
renewable energy generation decreases, iSwitch intelligently shifts the computing load
from one energy source to another to achieve best load power matching. We use virtual
machine (VM) live migration to implement iSwitch since it is the most convenient way to
perform load power shifting in a virtualized computing environment.
Note that in this work we consider server clusters that have reserved computing
capacity. In other words, in Figure 4-5, iSwitch does not require increasing the number
of servers to meet the workload surge. In addition, when power emergency happens,
we still use backup energy storage to temporarily support the load. According to Fan et
al. [43], a typical server clusters (not under-provisioned) can spend more than 80% of
the time within 80% of their peak power, and 98% of the time within 90% of their peak
power. Therefore, the chance of workload triggered emergency is small. In this study,
we assume that the number of renewable energy powered servers is less than 40% of
the overall deployed machines since a data center typically consumes about 60% of its
actual peak power [43]. In this case, even if the wind power is extremely low, the utility
grid can still take over most of the load power demand.
46
Figure 4-5. The iSwitch load power balancing scenario. The server clusters have reserved computing capacity and are powered by hybrid power supplies (i.e., wind turbine + utility grid).
To handle the time-varying, intermittent renewable power, iSwitch dynamically
allocates and de-allocates (i.e., “switches”) the renewable energy powered server load.
The supply/load variability makes the switch tuning challenging since the control should
1) globally respect the time-varying renewable budget and 2) locally avoid any power
failure induced by load fluctuation. To this end, iSwitch uses a hierarchical switching
control scheme, which can be easily incorporated into existing hierarchical power
management methods such as [44].
Facility level. Figure 4-6 shows a global view of iSwitch control mechanism.
The switching operation is controlled by a central switch controller (CSC), which
communicates with a central power controller (a typical facility-level data center power
controller), a switch scheduler and multiple cluster level switch controllers. CSC
performs switch tuning based on the discrepancy between the load power consumption
and the RES budget. Whenever needed, switching operations are scheduled by the
switch scheduler, which stores profiling information for each server load.
47
Figure 4-6. The iSwitch global (facility-level) power management.
Figure 4-7. The iSwitch local (rack-level) power management.
Cluster level. The switching allocation is assigned to computing nodes via
cluster-level switching controllers, which are counterparts to PDU-level power
controllers. It collects switching outcomes (i.e., the number of switching operations
accomplished/failed) of each computing node and feeds the information to the CSC for
switching scheduler updates.
Rack level. As shown in Figure 4-7, a rack-level switch controller executes
power supply switching and sends the execution outcomes to the CSC via a cluster-
level switch controller. It also interacts with the rack-level power controller throughout
the switching process to avoid any brownout. For example, whenever the power
consumption of a server rack reaches its local renewable power budget, the power
controller will signal the rack-level switch controller to throttle the switching activities.
4.3 Optimizing Load Tuning Activities
The switching scheduler is the key architecture for iSwitch, as shown in Figure
4-8. It monitors the power provisioning status (i.e. powered by renewable energy or
utility grid) of each server load (i.e., VMs). All the running loads within each cluster are
indexed consecutively in a switch allocation buffer (SAB). A switching history table is
used to store the switching frequency for each virtual machine. An optimizer computes
the optimal switching assignment and a tracking module initiates the switching process.
48
Figure 4-8. Structure of the iSwitch scheduler.
To make load tuning decisions, iSwitch scheduler needs profiling information
such as server utilization data from the load history table. The central power controller
invokes scheduling activities in response to variations in renewable power supply and
load power demand. Whenever necessary, the scheduler sends a sequence of load
switching commands to the central switch controller for execution.
4.3.1 Lazy Power Supply Tracking
The first idea of iSwitch is to avoid tracking the severely fluctuant renewable
power in Region-II. We call it lazy tracking because the module only harvests the
relatively stable renewable energy generation. Note that iSwitch carefully distributes the
switching activities across all the loads evenly to avoid local traffic jam.
Lazy tracking. At each fine-grained interval, when switching is triggered by the
CSC, an estimated switching assignment will be sent to the scheduler for calculating the
switch operation balance (e.g., estimated assignment minus the baseline). If the switch
balance indicates a reduced number of servers to RES connection, the scheduler
signals the CSC to schedule the estimated assignment to avoid brownout. On the other
hand, if the switching balance suggests an increased number of servers to RES
connection (e.g., due to temporally decreased load or increased supply), the scheduler
will signal CSC only if the balance is larger than a preset threshold (e.g., 10% of the
baseline). In this case, we ignore the high-frequency switching surge which brings little
benefit on renewable energy utilization but leads to excessive load migration.
49
LRU distribution. Within each cluster, iSwitch allocates switch operations with
least recently used (LRU) method, which avoids aggressively tuning a small set of
computing nodes. Note that a naive switching allocation can result in unbalanced
switching allocation. In Figure 4-9 we show the switching distribution as a result of
round-robin scheduling. The average switching frequency is 200 times per day per VM.
A small group of VMs receives up to 400 times per day. As a result, some racks may
incur more performance penalty due to high communication traffic.
To implement LRU, iSwitch uses the switch frequency record stored in the
switching history table. The operation of iSwitch scheduler relies on the load history
record of the previous control period. This record can be implemented using a round-
robin database (circular buffer) with constant storage occupation over time.
4.3.2 Dynamic Power Demand Smoothing
Optimizing the supply-side fluctuation alone cannot achieve significant overhead
mitigation. To this end, iSwitch leverages the heterogeneity of server clusters to
minimize load fluctuation-induced overhead in power management of region III.
Figure 4-10 illustrates the switch management timeline of iSwitch. The controller
re-shuffles the renewable energy powered servers at a coarse-grained time interval R
(e.g., 15 minutes as the default value in our experiments). During each re-shuffling
interval, the average load utilization is recorded in a fine-grained time interval (e.g., 1
minute) and is used to predict the load for the next period. Upon rescheduling, the
optimizer in the iSwitch scheduler updates the baseline switch operations of each server
cluster in SAB with the goal of mitigating the likelihood of severe load power fluctuation
in the next control period. Each switch tuning invoked by the CSC will be assigned
based on the updated SAB.
50
Figure 4-9. The histogram of switching frequency using round-robin virtual machine selection during load migration.
Figure 4-10. Optimization timeline of iSwitch. Lazy tracking is fine-grained
and demand smoothing is coarse-grained.
At the beginning of each control period, iSwitch recalculates the optimal switch
assignment. To simplify the problem, we assume the servers are logically divided into c
clusters and the load is balanced within each cluster (i.e., homogeneous utilization).
Let 𝑈𝑖 = [ 𝑢𝑖1 𝑢𝑖2… 𝑢1𝑐] denotes the average utilization of each cluster at
time stamp 𝑖. The utilization history record for the previous control period that consists of
𝑚 time stamps is:
𝑼 = [
𝑈1
𝑈2
⋮𝑈𝑚
] = [
𝑢11 𝑢12
𝑢21 𝑢22
… 𝑢1𝑐
𝑢11 𝑢2𝑐
⋮ ⋮𝑢𝑚1 𝑢𝑚2
⋱ ⋮𝑢11 𝑢𝑚𝑐
] (Equation 4-2)
Assuming a total number of 𝑁 virtual machines is to be connected to the
renewable power supply in the next control period. The migration decision for the next
control period is 𝑺 = [ 𝑠1 𝑠2 … 𝑠𝑘 … 𝑠𝑐], where 𝑠𝑘 is the number of VMs selected
to be tuned for cluster 𝑘. To reduce unnecessary load tuning in the future, we want the
51
aggregate power consumption of the selected VMs to have small oscillations in the next
control period. In other words, the standard deviation of the aggregate utilization should
be minimized. The aggregate utilization is given by:
[𝑎𝑖𝑗]𝑚×1
= 𝑼 × 𝑺𝑇 (Equation 4-3)
The standard deviation of the expected utilization in the next control period can
be calculated as:
𝛿 = √1
𝑚∑ (𝑎𝑖 − 𝜇)2𝑚
𝑖=1 = √1
𝑚∑ (𝑎𝑖)2𝑚
𝑖=1 − (1
𝑚∑ 𝑎𝑖
𝑚𝑖=1 )2 (Equation 4-4)
𝑎𝑖 = ∑ 𝑢𝑖𝑘𝑠𝑘𝑐𝑘=1 (Equation 4-5)
In Equation-4, 𝑎𝑖 is the aggregate utilization of renewable energy powered load
and 𝜇 is the mean utilization during the past control period 𝑅. The re-shuffling problem is
therefore formulated as:
Objective: min {1
𝑚∑ (𝑎𝑖)2𝑚
𝑖=1 − (1
𝑚∑ 𝑎𝑖
𝑚𝑖=1 )2} (Equation 4-6)
Constraints: ∑ 𝑠𝑘 ≤ 𝑁 (Equation 4-7)
We solve the above non-linear minimization problem with simulated annealing
(SA). Given the utilization history records, our SA solver is capable of finding the
desired global extreme very fast. Note that the renewable power supply fluctuation
typically occurs on a coarse-grained time interval (several minutes). As a result, the
execution time (several seconds in our experiments) of SA solver does not affect the
optimization effectiveness. At the beginning of the re-shuffling period, the switching
operations are assigned in proportion to the number of servers in the cluster. The SA
solver iteratively generates a stochastic perturbation for the switching assignment and
checks whether or not the optimum solution is reached.
52
4.4 Experimental Methodology
We evaluate our design with trace-driven simulation. We have developed a
framework that simulates dynamic load tuning and hierarchical power control in
renewable energy powered data centers. For each scheduled job requests, we calculate
its contribution to the overall data center power consumption based on the number of
nodes requested and the job’s specific resource utilization statistics. Our framework
takes realistic wind energy generation traces as supply-side input.
We assume a raised floor data center consisting of 4,800 servers and the server
we modeled resembles an HP ProLiant DL 360 G6. The peak and idle power of the
modeled server are 186W and 62W respectively. We convert the server utilization
traces to its power consumption using the published SPECPower results [45], which
have been widely used for data center power evaluation [37] [38]. Since the
SPECPower results only reflect the server power at intervals of 10% utilization, we use
linear interpolation to approximate the power across different load levels.
Our heterogeneous data center configuration is based on an academic high-
performance computing (HPC) center in the department of electrical and computer
engineering at the University of Florida, which hosts over 600 servers. The HPC center
has five major clusters (C-I~C-V) with different service targets and loads, as detailed in
Table 4-1. Those clusters have 20 to 400 computing nodes and their average utilization
ranges from 25% to more than 90%. Since we have limited access to industrial data
center traces, we collect real-world workload logs from a well-established online
repository [46]. The workload logs provide information such as job arrival times, start
time, completion time, and number of requested nodes.
53
Table 4-1. Configurations of heterogeneous clusters.
Cluster ID % of Overall Deployed Servers Average Cluster Loading
C-I 5% 97% C-II 63% 60%
C-III 17% 57%
C-IV 3% 54%
C-V 12% 25%
Table 4-2. The evaluated data center workload sets. Workload Set Workload Trace Combination
Mix-High “HPC2N” + ”LANL CM5” + ”LPC EGEE” + “SDSC BLUE” + “LLNL Thunder”
Mix-Low “DAS- fs0” + “DAS2-fs1” + “DAS2-fs2” + “DAS2-fs3” + “DAS2-fs4”
Mix-Stable “HPC2N” + “KTH SP2” + “LANL CM5” + “DAS2-fs0” + “SDSC BLUE”
Mix-Bursty “DAS2-fs2” + “DAS2-fs3” + “DAS2-fs4” + ”LPC EGEE” + “OSC Cluster”
Mix-Rand “LLNL Thunder” + “OSC Cluster” + ”LPC EGEE” + ”LANL CM5” + “KTH SP2”
Dept-HPC Five traces collected from departmental high-performance computing center
Table 4-2 summarizes our evaluated workload trace combinations. In additional
to the load traces from our local HPC center (i.e., Dept-HPC), we also build various
workload mixes to mimic co-located clusters. Each workload set in Table 4-2 consists of
five traces which run on the five clusters (C-I to C-V) respectively. We characterize
workload traces based on their average job size and runtime. In Table 4-2, Mix-High
includes traces that have larger job size (resulting in >30% average data center
utilization) and Mix-Low contains traces that have small job size (resulting in <10%
utilization). On the other hand, Mix-Stable consists of five traces that feature relatively
longer job runtime and Mix-Bursty consists of traces that have very short job runtime.
We use wind power data traces from The Wind Integration Datasets [47]. In
Table 4-3, capacity factor is the ratio of the actual wind turbine output to the theoretical
peak output. We have selected a group of traces with various capacity factor values.
Typically, the capacity factor is 30% [35]. The total installed wind power capacity in this
study equals to the nameplate power of data centers. The actual power budget is
therefore only affected by the capacity factor.
54
In Table 4-4 we list our evaluated power management schemes, Utility and
Battery are two conventional schemes which do not adjust server load to match the
renewable power budget. Green is the most sustainable design that only uses wind
energy to power server clusters. Tracking represents emerging design approaches
which adjusts server load to actively track every joule of renewable energy generation
(i.e., place as much workload as possible on green clusters) [10] [22].
Table 4-3. The evaluated wind power supply traces. Traces Wind Energy Potential Locations (ID) Capacity Factor Power Density
W1 Low California (9250) 15% 195 W/m2 W2 Medium Arizona (6107) 25% 338 W/m2 W3 Typical Colorado (10563) 35% 581 W/m2 W4 High Texas (1360) 45% 708 W/m2
Table 4-4. The evaluated server cluster power management schemes. Abbr. Design Philosophy Power Tracking Stored Energy Load Deferment
Utility Fully relies on utility grid No No No Battery Relies on battery No Yes No Green 100% renewable energy Yes No Yes Tracking Aggressive tracking Yes Yes No iSwitch Low overhead and latency Yes Yes No
4.5 Evaluation Results
We first characterize the impact of supply/load power variability on data center
load matching using homogeneous workload traces. Frequent load matching activities
result in operational overhead, which is our primary concern in the design of renewable
energy powered computing systems. In a virtualized environment, iSwitch could
effectively reduce the VM migration rate and help to save data center network
bandwidth significantly. The data migration traffic is calculated at rack-level. Each VM
live migration transfers approximately the size of the VM’s memory between hosts [48].
We assume a VM memory size of 1.7GB in our calculation, which is the default memory
size of Amazon EC2 standard instance.
55
Figure 4-11. The average network traffic. All the results are normalized to Tracking.
Figure 4-12. The peak network traffic. All the results are normalized to Tracking.
Figure 4-11 shows the average communication traffic across various workload
configurations and wind power supply levels. All the results are normalized to Tracking.
We do not show the results of Green because it has the same power tracking frequency
as Tracking. As can be seen, on average, iSwitch could reduce 75% of the rack-level
traffic and therefore significantly releases the network bandwidth burden. The results
are even more impressive for peak traffic hours when the renewable energy fluctuates
severely. In Figure 4-12 we calculate the communication traffic during the top 1% high-
traffic hours. Because iSwitch puts a limit on the power tracking activities during
fluctuant supply period, it shows only 5% network traffic compared with Tracking.
Another advantage of iSwitch is that it reduces the migration frequency of each
VM instance and thereby improves the job turnaround time. We calculate an execution
time increase (ETI) for each job to evaluate the impact of load shedding and workload
deferment on job execution time. Even for utility-connected systems such as iSwitch
56
and Tracking, ETI exists due to the data migration. For example, the time needed for a
1GB VM migration takes about 20 seconds to complete [49].
We use ETI per hour (EPH) to evaluate the performance overhead. For example,
an EPH of 10 means each individual VM instance experiences 10 seconds waiting time
per hour on average. Figure 4-13 shows the average EPH across the entire data center
servers. The average EPH of iSwitch is about 30 seconds per hour while the average
EPH of Tracking reaches 126 seconds per hour. The average EPH of Green, however,
is about 1500 seconds per hour – 50 times that of iSwitch. Therefore, waiting for
renewable energy to resume (e.g. Green) should be the last resort.
We evaluate the renewable energy utilization (REU) of data centers with different
wind power provisioning capacities and workload behaviors. The REU is defined as the
amount of wind energy that is actually utilized by the load divided by the total amount of
wind energy generation. A higher REU indicates better supply/demand coordination,
which reduces on-site energy storage capacity, improves return-on-investment (ROI)
and data center sustainability, and eases the initial infrastructure planning.
While iSwitch uses a lazy power tracking scheme, it does not sacrifice energy
utilization significantly. As shown in Figure 4-14, iSwitch can achieve an average
renewable energy utilization of 94% – higher than Green (92%) but lower than Tracking
(98%). The reason why Tracking outperforms iSwitch on energy utilization is that
Tracking tracks every joule of wind energy generation aggressively. Note that a 4%
decrease in energy utilization does not mean that iSwitch is less preferable in our study;
iSwitch significantly reduces network traffic and improves the performance by 4X. In
contrast to Tracking, iSwitch trades off energy utilization for better turnaround time.
57
Figure 4-13. The average execution time increase (ETI) of job requests.
Figure 4-14. The normalized overall wind energy utilization.
4.6 Summary of iSwitch Design
Matching the variable load power consumption to the intermittent power supply
appropriately is the crux of designing a renewable energy powered data center.
Conventional workload-driven power management has less adaptivity to the power
supply variation while existing power-driven system control schemes are not aware that
they are experiencing unnecessary and redundant tuning activities. As a result, we
either miss the opportunity of utilizing precious renewable energy or incur significant
load tuning overhead and performance degradation.
This chapter proposes iSwitch, a renewable energy-driven power tuning scheme
that addresses a two-fold challenge: the first is to manage intermittent renewable power
without power throttling; the second is to ensure high energy utilization with minimized
load matching activities. To the best of our knowledge, this is the first work that digs into
renewable power variation characteristics and introduces a supply/load cooperative
optimization scheme that minimizes power management overhead.
58
CHAPTER 5 DATA CENTER LEVEL GREEN POWER MANAGEMENT:
TOWARDS LOAD-FOLLOWING BASED DESIGN
Supply-tracking (a.k.a. supply-following) can eliminate supply-demand power
mismatches in renewable energy powered computing systems, as demonstrated in
previous SolarCore design and iSwitch design. In this chapter we discuss a radically
different design approach, which allows near-optimal performance of computing
systems on green energy resources. The key idea is to put emphasis on enabling
renewable energy supply to follow power demand, rather than forcing server systems to
track the variable power budget. To achieve this, we leverage distributed generation
initiated by the smart grid technology [50].
Distributed generation (DG) refers to a variety of small, modular electric
generators near the point of use. In recent years, DG has gained tremendous interest
as an alternative source of power for IT industry. According to the U.S. Environmental
Protection Agency (EPA), using DG in data center design could achieve great energy
savings, significant environmental benefits, and high power reliability [51].
As shown in Figure 5-1, DG system encompasses a wide range of green energy
technologies, such as solar panels, wind turbines, fuel cells, and bio-fuel based gas
turbines. While solar/wind power heavily depends on environmental condition, the
outputs of fuel cells and gas turbines are tunable. They can provide a key supporting
service called load following [52], which refers to the use of online generation equipment
to track the changes in customer loads. Therefore, one can take advantage of the load
following capabilities of these tunable DG systems to meet the time-varying server
power demand. Such design is non-trivial because it enables computing system to run
on renewable energy sources without compromising workload performance.
59
Figure 5-1. Distributed generation (DG) powered data centers. Some of the tunable
DG systems can be controlled to follow the load power demand.
Figure 5-2. Load following scenario. Fine-grained server power demand fluctuation
hinders efficient load following on distributed generators.
When employing distributed generation to build a better data center power
provisioning architecture, challenges arise due to the unpredictable and fluctuating data
center load. Figure 5-2 shows typical load following scenario that tracks customer load
every 30 minutes. As can be seen, DG systems mainly provide coarse-grained load
demand following due to limited response speed of onsite generators. To handle the
moment-to-moment load power demand, DG typically relies on large on-site energy
storage elements [52]. Such design not only increase the TCO (due to storage cost and
maintenance), but also incurs up to 25% roundtrip energy loss.
We propose data center power demand shaping (PDS), a novel power
management approach that enables high-performance low-overhead data center
operation on pure renewable energy sources. The novelty of PDS is two-fold. First, PDS
intelligently trims data center load power and enables DG systems to follow the power
60
demand efficiently. Second, PDS features two adaptive load tuning schemes that could
boost data center performance and enable near-oracle operation during power demand
trimming process. As a cross-layer power optimization scheme, our power management
module resides between front-end distributed generation and back-end computing
facilities to provide a coordinated tuning between the supply and load.
5.1 The Load Following Challenge
Different from conventional bulk grid which has large amount of capacity inertia,
distributed generators does not have reserved capacity [51]. The supply and storage of
energy must be planned carefully to ensure instantaneous energy balance [50].
Most energy storage devices have very fast response speed that could release
power almost immediately. As a result, they are widely used to handle moment-to-
moment load oscillation and disturbances, which is referred to as regulation [52]. In
contrast, gas turbines and fuel cells are typically too slow to meet the load power
variation since the change of the engine speed or the chemical reaction in the fuel
requires time. Therefore, they are used to track the intra- and inter-hour changes in
customer loads, which is referred to as load following [52]. Although there is no strict
rule to define the temporal boundary between regulation and load following, typically
load following occurs every 10~15 minutes or more [52]. It is not economically feasible
to frequently adjust the power output of distributed generators due to the increased
performance cost and decreased fuel utilization efficiency.
The energy balance issue arises due to the fluctuating load in data centers and
other computerized environments. Dynamic power tuning techniques (e.g., DVFS),
frequent on/off power cycles, stochastic user requests, and data migration activities can
cause varying degree of load power variation. Since distributed generators are generally
61
placed near or at the point of energy consumption, they are often exposed to the full
fluctuation of local IT loads rather than experiencing the averaging effect seen by larger,
centralized power system. As a result, energy balancing becomes rather difficult in
distributed generation powered data centers.
5.2 Distributed Generation Powered Data Centers
Conventional design always focus on managing data center power demand and
overlooks the capability of power supplies. In contrast, our system has two tuning
knobs: the onsite generation level and the server power demand.
We adopt typical microgrid power provisioning hierarchy for managing various
distributed energy resources, as shown in Figure 5-3. The design consists of a group of
radical feeders to provide power to electrical loads. Microsources are connected to the
feeder through circuit breakers and coordinated by the microgrid energy management
module (EMM), which provides real-time monitoring and autonomic control of the power
generation [53]. Although microgrid can connect to the utility through a single point
called point of common coupling (PCC), we focus on standalone microgrid due to
sustainability and cost-effectiveness reasons. Such islanded mode is also a remarkable
feature of microgrid to avoid power quality issue in the main grid [53].
Figure 5-3. Distributed generation powered data centers. A power demand controller is used to coordinate onsite generators, batteries, and data center loads.
62
The connected electrical loads may be critical or non-critical. HPC servers are
typically sensitive and mission-critical loads that require stringent power quality and
sufficient power budget. It is preferable to assign these loads with controllable and
stable microsources such as gas turbines and fuel cells.
We use a power demand controller (PDC) to manage green data centers. It
keeps monitoring the job running status and the overall power demand. It also
coordinates with the job scheduler for the purpose of better computing resource
allocation. Meanwhile, it also communicates with the microgrid EMM to obtain current
DG levels and dynamically inform the EMM for necessary adjustment.
A microgrid system typically uses frequency droop control to monitor and adjust
the on-site power generation [54]. This scheme uses a small change in the power bus
frequency to determine the amount of power that each generator should put into the
power bus. In addition, the microgrid also includes energy storage devices to provide
necessary startup time for generators’ output to ramp up when system changes. In this
study we leverage uninterruptible power supplies (UPS) system to provide necessary
stored energy. Note that although our proposed technique also applies to centralized
UPS systems, we choose emerging distributed UPS architecture which is known to
have higher energy efficiency and better scalability [17].
5.2.1 Power Demand Shaping Mechanism
While most of the related work on green data center focuses on managing the
load, we have identified a non-conventional supply-side opportunity. We propose power
demand shaping (PDS) technique to better utilize onsite distributed generation in
modern HPC data centers. The idea is to aptly influence data center power demand to
match the planned power usage curve that yields high load following effectiveness.
63
Although one can transform the IT power demand into many different shapes, we
found that a square-wave-like demand curve is the most convenient, as shown in Figure
5-4. In this case, the distributed generation system only needs to increase (ramp up) its
generation at each end of the load tuning cycle (e.g., every 15 minutes) to meet the new
power budget goal. Since the load power demand becomes less bursty and fluctuating
after tuning, there is no need to install bulky energy storage devices onsite. In this work
we mainly leverage the UPS to handle the ramping period and the temporary power
discrepancy between supply and load. By eliminating the unnecessary energy storage
capacity, our design also becomes more cost-effective and sustainable.
Figure 5-4 demonstrates our power demand shaping mechanism in terms of
power supply behavior and load performance status. The proposed scheme uses a
three-step control process in which data center load is adaptively managed and the
distributed energy is carefully scheduled.
Figure 5-4. Power demand shaping demonstration. The idea is to gracefully
influence the data cater power demand to generate a planned, DG-friendly power usage curve.
64
Step 1: Maintain constant power demand. During each control period, the
power demand controller uses power capping to maintain a constant power demand.
Assume the original load power demand (i.e., the expected power consumption without
any load tuning) is D, we scale down the load performance whenever D exceeds current
power demand goal. If the power demand is below the power supply (rarely occurs
since we under-provision the system), we store the excess power in UPS batteries.
Depending on the system’s hardware characteristics and BIOS support, the
actual performance scaling policy may vary. In this work we take advantage of
performance states (a.k.a. P-states) to dynamically adjust CPU frequency to match the
required budget. This is a technique to manage power by running the processor at a
less-than-maximum clock frequency. We do not use voltage scaling because it may not
yield notable cubic power saving in the real measurement [40]. It has been shown that
the power-to-frequency curve for both DFS and DVFS can be approximated as a linear
function [40]. As a result, we use only frequency scaling for simplicity.
Step 2: Re-schedule DG generation. In contrast to existing power capping
scenario, PDS deals with time-varying power capping budget. At the end of each control
period, the controller will inform the distributed generation EMM to update the
generation level. To calculate the amount of generation adjustment, PDS calculates the
mean value of the original power demand in the past control period. The actual DG
power adjustment is selected from the mean value and current demand measurement,
depending on whichever is greater. After that, the distributed generation will gradually
increase its output and we assume 1-min power ramping for the generators. During this
period, we leverage onsite UPS storage to provide necessary power support.
65
Step 3: Adjust node performance. At the beginning of each new period, we
update the node performance level using the newly scheduled energy. There are
several different scaling schemes. One can improve the frequency of every running
node to achieve a fair performance tuning. In addition, we can use a priority-based
scheme. It associates each node with a priority and makes a scheduling choice based
on the priorities. For example, a node with the highest priority gets highest frequency.
In Figure 5-5 we show the pseudo code of power demand shaping. The controller
monitors the data center load every tick (i.e., every second in our design) and adjust
node frequency based on the discrepancy between the budget and the original demand.
At the end of each cycle TL, the controller adjusts the budget for a new demand goal.
Definition: TL: Load following interval; D: Maximal power demand of
current load; P: Mean demand in the last control interval
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
for Each load tuning timestamp tick
if tick % TL == 0 // update power budget
Budget ← max(P, D);
else // maintain constant power demand
D ← load power measurement
if load demand changes
ΔPLF ← (Budget – D)
Adjust load performance based on ΔPLF
end if
end if
end for
Figure 5-5. Power demand shaping algorithm.
5.2.1 Adaptive Load Performance Scaling
Strict power demand shaping facilitate load following but comes with
performance degradation. To this end, we also propose adaptive load performance
scaling that helps PDS to improve job performance while maintaining high load following
effectiveness. The key idea is to use a relaxed power demand shaping policy [24].
66
In addition to the scheduled generation adjustment, the controller also assigns a
bonus power budget for two reasons. First, after a period of time, the UPS requires
recharge to retain its full capacity. In this case, the PDS will assign additional generation
in the next period based on the required charging power. Second, PDS features an
adaptive load performance scaling scheme which intelligently leverages DG energy and
UPS energy for boosting the load performance.
DGR boost. The first optimization scheme helps to fine-tune the distributed
generation level [24]. Since we have no oracle knowledge of the incoming task,
adjusting DG generation based on historical average unavoidably incurs prediction
error. All the active jobs under this circumstance suffer low processor speed and the
negative impact can last as long as the load tuning interval. To solve this problem, our
DGR Boost optimization dynamically monitors the normalized job execution progress
and the overall system throughput, and leverages the over-clocking capabilities of
modern processor nodes to minimize the job delay. To enable DGR Boost, the PDS
controller needs to dynamically monitor the execution status of each active job.
UPS boost. This scheme fine-tunes intra-cycle load power [24]. The main idea is
that for short-running jobs (jobs that finish within one power demand control interval)
that cannot gain the benefits of DGR Boost, we can leverage the UPS stored energy to
provide additional power support and avoid significant performance degradation. In
typical HPC data centers, users are required to submit their job runtime estimations to
enable backfilling, which can help maximize cluster utilization and throughput. In this
study we use job runtime to sort out short jobs. The short jobs here are defined as
tasks that will complete before current load following cycle ends.
67
5.3 Experimental Methodology
We develop a HPC data center simulation infrastructure that takes real-world
workload traces as input. Our discrete-event simulator puts each user’s request into a
queue and waits to grant allocation if computing nodes are available. Each job request
in the trace has exclusive access to its granted nodes for a bounded duration. Such
trace-driven discrete-event simulation has been adopted by several prior studies on
investigating data center behaviors and facility-level design effectiveness [37] [38].
We simulate the power behavior of distributed generation system on per-second
time scale which is in tune with our data center simulator. The distributed generator
adjusts its output on a coarse-grained interval (10-minute as the default value) and
batteries respond to the fluctuating computing load. We adopt microsource optimization
model HOMER in our simulator [55]. HOMER simulates both conventional and
renewable energy technologies and evaluates the technical feasibility of a large number
of power system options. We assume a 1-minute constant duration for the distributed
generation to ramp up when perform load following. The battery cycle life is set to be
10,000 times and we calculate the capacity degradation of battery cell based on its
discharging profile which includes detailed information of each discharging event.
We still use real-world workload traces from [46]. We use five key task
parameters of each trace file: job arrival time, job start time, job completion time,
requested duration, and job size in number of granted computing nodes. As shown in
Table 5-1, we select six 1-week workload traces that have different mean utilization
level and mean job runtime. Our simulator uses batch scheduling and the job
scheduling policy is first come first serve (FCFS). We examine the data center load and
distributed generation budget at each fine-grained simulated timestamp.
68
Table 5-1. The evaluated HPC data center workload traces.
Trace Description Avg. Load
Mean Inter-arrival Time
Avg. Job Run Time
LANL Los Alamos Lab’s 1024-node machine 56% 4.9 min 31.6 min SDSC San Diego Supercomputer Center’s Blue Horizon 60% 3.7 min 33.0 min Atlas Lawrence Livermore Lab’s 9216-CPU cluster 46% 10.6 min 36.8 min Seth A 120-node European production system 85% 20.5 min 6.2 h MCNG A 14-cluster data center, 806 processors in total 89% 2.2 min 11.1 h RICC A massively parallel cluster with 8k+ processors 52% 0.9 min 16.6 h
Table 5-2. The evaluated data center power management schemes.
Schemes Descriptions
Oracle Ideal power provisioning with no performance overhead LF Existing load following based design ST Existing supply tracking based design PDS-s PDS without optimization ( i.e., uses strict power budget ) PDS-r PDS + adaptive load tuning (DGR Boost & UPS Boost)
5.4 Evaluation Results
In this section we discuss the benefits of applying power demand shaping to
distributed generation powered data centers. In Table 5-2, Oracle is an ideal design that
has a priori knowledge of load patterns and could always meet the fluctuating data
center power demand. It represents the ideal energy balance scenario that one can
achieve with renewable energy resources. Different from Oracle, LF is a conventional
load following based scheme that has heavy reliance on energy storage. ST represents
existing power supply driven design that aims at managing the computational workload
to match the renewable energy supply [10] [12]. PDS-s is our proposed power demand
shaping mechanism without optimization. PDS-r is the relaxed power demand shaping
that features adaptive workload-aware performance scaling.
Performance is one of the key driven forces of our power demand shaping
technique. Although tracking renewable power budget has shown great success on
reducing IT carbon footprint, it cannot ensure performance and sustainability
69
simultaneously. Table 5-3 evaluates the performance degradation of supply tracking
based design under different wind energy intensity. We have scaled the wind power
traces so that the average wind power budget equals the average data center power
demand. When renewable energy generation drops or intermittently unavailable, we
scale down the node frequency and defers job execution if it is necessary.
Table 5-3. Mean job turnaround time of tracking based design (ST) under different renewable generation levels. All the results are normalized to Oracle.
LAN SDSC Atlas HPC MCNG RICC
200 W/m2 2.92 1.60 1.64 7.25 1.31 1.24 400 W/m2 2.10 1.51 1.66 9.27 1.12 1.29 600 W/m2 2.89 1.41 1.37 5.91 1.08 1.10 800 W/m2 5.27 1.34 1.30 7.00 1.06 1.08
The results in Table 5-3 show that data centers designed to track the variable
renewable power supply could incur up to 8X job runtime increase. The geometric mean
value of turnaround time increase is 59%. It also shows that workload performance
heavily depends on renewable power variation behavior and data center load pattern.
We have observed that occasionally the renewable energy variation pattern is totally
uncorrelated with the load power demand. There is no guaranteed performance lower
bound since both the user behavior and environment conditions are stochastic.
In contrast to supply tracking based design, load following fundamentally
changes this situation. Figure 5-6 shows the mean job turnaround time under varying
load following interval. All the results are normalized to Oracle, which has no
performance scaling and maintains full-speed server operation with sufficient power
budget. The average execution time increase (ETI) of PDS-s is 8.0%. Our optimized
scheme, PDS-r, shows 1.2% mean performance degradation compared to Oracle. This
means that power demand shaping with adaptive load tuning (i.e., DGR Boost and UPS
Boost) could yield a 37% improvement over existing supply-tracking based design.
70
Figure 5-6. Average job turnaround time under varying load following interval. All the
results are normalized to Oracle.
A B
Figure 5-7. Execution time of the worst 5% jobs. A) Results of PDS-s. B) Results of PDS-r. All the results are normalized to Oracle.
In addition to the mean performance statistics, Figure 5-7 further investigates the
worst-case scenario in terms of the mean job turnaround time of the worst 5% jobs. We
observe that PDS-s and PDS-r shows 47% and 9.6% performance degradation,
respectively. Therefore, PDS-r shows better performance guarantee in terms of the
worst-case performance degradation.
Conventional green energy-aware scheduling schemes normally rely on utility
grid. To achieve the same performance as PDS-r, these designs have to increase their
reliance on the utility grid, as shown in Figure 5-8. The percentage of time that utility
power is used is between 30% and 70%. The geometric mean value across all
workloads and renewable generation levels is 51%. Heavy reliance on utility means
increased carbon footprint and utility power cost, which goes against the original
intention of sustainability.
71
Figure 5-8. Dependence on utility power. To compete with PDS-r, even state-
of-the-art designs have to heavily rely on the utility power.
Compared to conventional load following, PDS results in much lower UPS stress.
In Figure 5-9B, both PDS-s and PDS-r shows less than 3% battery lifetime degradation.
Although PDS-r explores additional stored energy to perform performance boost, its
impact on UPS system is very similar to that of PDS-s. Note that for PDS, the battery
wear out problem is alleviated when we increase the interval of LF cycles. This trend is
different from what we observed in LF. The reason is that batteries are mainly used to
provide power shortfall during generator’s ramp-up period in PDS based system. When
we increase the LF intervals, we also lower the frequency of generator ramp-up, the
dominant factor of battery discharging.
Another advantage of PDS is that it allows one to use conventional centralized
UPS system to assist power demand shaping. The drawbacks of centralized UPS
system is that it cannot provide a fraction amount of energy – all the data center load
has to be switched between the main supply and UPS. Consequently, during each UPS
ride-through, UPS batteries have to experience much higher discharge current. The
immediate result of this is significantly decreased UPS lifetime. As shown in Figure 5-
10, LF in this case results in 35% lifetime degradation on average. The lifetime
degradation of PDS-s and PDS-r is 6% and 7%, respectively. Therefore, we believe
PDS maintains acceptable overhead and can be applied to centralized UPS systems.
72
A B
Figure 5-9. Battery lifetime degradation comparison (with distributed UPS). A) Load following based design. B) PDS design.
Figure 5-10. Battery life degradation comparison (with centralized UPS).
5.5 Summary of PDS Control
Distributed generation (DG) systems are gaining popularity and their installed
capacity is projected to grow at a faster pace. We expect this trend to continue, as the
energy crisis and environmental problem become increasingly crucial to our planet.
In this work we investigate how the incoming smart grid incentive would impact
the design and optimization of data centers. We propose distributed generation
powered data center that leverages the load following capability of onsite renewable
generation to achieve low carbon footprint without compromising performance. The
unique feature of load following based design makes this work a step forward toward
the goal of incorporating a variety of clean energy resources into computing systems.
We have developed novel power demand shaping technique (PDS) to improve the load
following efficiency in data centers while boosting the workload performance with two
adaptive load tuning schemes: DGR Boost and UPS Boost. Overall, PDS could achieve
98.8% performance of an ideal oracle design.
73
CHAPTER 6 ENABLING DATA CENTER SERVERS TO SCALE OUT SUSTAINABLY
As cloud applications proliferate and data-processing demands increase, server
resources must grow to unleash the performance of emerging workloads that scale well
with large number of compute nodes. Over time, the constant influx of server resources
in data centers will eventually become power-constrained. According to a recent
industry survey from the Uptime Institute [56], 30% of the enterprise data center
managers expected to run out of power capacity within 12 months.
While leasing collocated racks and deploying cloud services fit small server
clusters on a budget, this study mainly focuses on large-scale enterprise data centers
and approaches that aim at improving computing capability and capacity. Server
consolidation is mature techniques that can free up power capacity. However, even
consolidated servers have to limit their performance using either software-based (i.e.,
virtual CPU allocation) or hardware-based (i.e., DVFS) control knobs to avoid tripping
circuit-breakers and causing costly down-time.
Upgrading power systems is a radical solution that allows one to add more
servers, racks, and even onsite containers. However, like building a new data center,
re-sizing data center initial power capacity can be a great undertaking since
conventional centralized power provisioning scheme does not scale well. In a typical
data center, the power delivery path involves multiple power equipment across several
layers, as shown in Figure 6-1. Upgrading power infrastructure often requires re-design
of the entire power delivery path which is not only costly but also time-consuming.
Worse, utility power feeds are often at their capacity in some urban areas and additional
electricity access for data center is restricted.
74
In addition, modern scale-out data centers are not only power-constrained, but
also carbon-constrained. As server power demand increases, the associated carbon
footprint expansion poses significant challenges for scale-out data centers. It has been
shown that global greenhouse gas (GHG) emissions could easily exceed 100 million
metric tons per year if we keep using utility grid power that is generated from
conventional fossil fuel [57]. An emerging solution to the carbon reduction problem is to
leverage green energy sources to lower the environmental impact of computing
systems. Unfortunately, to the best of our knowledge, existing green energy integration
schemes often employ centralized power integration architecture and largely overlook
the modularity of renewable power supplies. As shown in Figure 6-1, facility-level
integration allows renewable power supplies to be synchronized to the utility grid.
However, the operation of such grid-connected renewable power system often relies on
the availability of utility power. Meanwhile, the centralized power integration not only
results in single-point of failure, but also makes future capacity expansion expensive.
Figure 6-1. Conventional data center power architecture is not ‘scale-out friendly’.
75
In this chapter we ask one important question: “Can we enable a power-
constrained data center to gracefully scale out while lowering the carbon
footprint with high efficiency?” Faced with the ever-growing computing demand, the
skyrocketing server power expenditures, and the looming global environmental crisis,
this question is of significant importance to data center operators who have to embrace
efficient power provisioning to survive economically and sustainably.
We propose Oasis, a unified power provisioning scheme that synergistically
integrates energy source control, power system monitoring, and architectural support for
power-aware computing. Oasis leverages modular renewable power integration and
distributed battery architecture to provide flexible power capacity increments. It allows
power-/carbon- constrained systems to stay on track with horizontal scaling and
computing capacity expansion.
We implement Oasis as a prototype research platform for exploring key design
tradeoffs in multi-source powered data center. Our first generation prototype is a micro
server rack (12U) that draws power from onsite solar panels, conventional utility power,
and local energy storage devices. These energy sources are coordinated by a micro-
controller based power control hub (PCH) built from scratch. We customize the PCH to
make it a rack-mounted, interactive system for easy installation and diagnosis.
We further propose Ozone, a power management scheme for Optimized Oasis
Operation (O3). Transcending the boundaries of traditional primary and secondary
power, Ozone creates a smart power source switching mechanism that enables server
system to deliver high performance while maintaining desired system efficiency and
reliability. Ozone is able to dynamically distill crucial runtime statistics of different power
76
sources to avoid unbalanced usage of different power sources. It could identify the most
suitable control strategies and adaptively adjusts the server speed via dynamic
frequency scaling to maximize the benefits of Oasis system.
To summarize, Oasis and Ozone constitute an interesting platform from many
perspectives: (1) it enables data centers to scale-out and facilitates partial green power
integration; (2) it links power supply management and server system control, and could
enable real-time power supply driven workload scheduling; (3) its power provisioning
architecture (i.e., hybrid power supplies + distributed control domain) could bring us
improved data center availability in many power failure scenarios; (4) its decentralized
multi-source power management could offer the flexibility of offering different green
computing services based on different customer expectations.
6.1 Oasis Scale-Out Model
We classify existing power capacity scale-out model as either utility power over-
provisioning or centralized power capacity expansion. In the first case, the utility grid
and data center power infrastructure is designed to support the maximal power the load
may ever draw. Although such design provides abundant power headroom, it inevitably
increases the carbon footprint. In the second case, the power delivery infrastructure is
provisioned for a certain level of anticipated future power drawn and the scaling out is
handled entirely by data center-level power capacity integration. However, installing
large-scale renewable power system often results in high expansion cost.
Oasis explores a “pay-as-you-grow” model for scale-out data-centers, which we
refer to as distributed green power increments model. In this model, the utility grid and
data center power delivery infrastructure are provisioned for a fixed level of load power
demand. When data center reaches its maximum capacity, however, we add renewable
77
power capacity by small increments in a distributed manner. This approach not only
provides carbon-free power capacity expansion to a power-constrained data center, but
also avoids over-committing its capitals. As shown in Figure 6-2, Oasis consists of a
number of green energy powered computing racks, called Oasis nodes. Each Oasis
node is attached to a power control hub (PCH), which further connects to renewable
power supply, power distribution unit (PDU), and distributed energy storage devices.
6.1.1 Modular Power Sources
Scale-out data centers need both modular standby power sources that provide
incremental backup power capacity and modular primary power sources that generate
additional electrical power. Today, distributed battery architecture is emerging to
improve data-center efficiency [17] [58]. It provides us a convenient way to add
additional backup power as computing demand increases. In addition, renewable power
supplies such as solar panels are usually modular and highly scalable in their capacity.
They are ideal power supplies that support carbon-free server scale-out.
Figure 6-2. Oasis power provisioning architecture. It leverages the modularity of both
battery systems and renewable power systems.
78
Figure 6-3 illustrates the distributed battery architectures in the Open Compute
Project [58] led by Facebook. In the coarse-grained integration scheme, a battery
cabinet populated with commodity lead-acid batteries is used to provide standby power
for a rack triplet. Each triplet consists of three column racks and each rack is further
divided into three power zones. The battery cabinet also includes several breakers,
quick fuse disconnects, sensors, and a high current DC bus bar. In the fine-grained
integration scheme, the battery cabinet is replaced by a high-density lithium-ion battery
backup unit (BBU) in each rack power zone. In both cases, battery provides 48V DC
backup power and could provide around 45 seconds of runtime at full load.
Figure 6-3. Distributed energy storage system (battery) emerge in
recent implementation of server clusters.
On the other hand, wind turbine and solar panel are both modular power
sources. Compared to wind turbines, solar panels can provide even smaller capacity
increments. In this study we look at solar panels that use micro-inverter [59] to provide
incremental power capacity. Conventionally, solar power system uses string inverters
which require several panels to be linked in series to feed one inverter. String inverters
79
are big, prone to failure from heat, and shows low efficiency. In contrast, micro-inverters
are smaller in size and can be built as an integrated part of the panel itself. In Figure 6-
2, each panel has its own micro-inverter; solar panels are further connected in parallel
to offer larger power capacity. Compared to centralized design, solar panel with micro-
inverter shows much better scalability, reliability and efficiency. The only disadvantage
is its relatively high cost per watt compared to centralized inverter. However, the
amortized cost can be lower than a centralized system (Section 6.5).
6.1.2 Distributed Integration
Oasis adds renewable power capacity at power distribution unit (PDU) level. As
shown in Figure 6-2, the PDU-level renewable energy integration allows us to scale out
server system on a per-rack basis. We do not use centralized power integration since it
does not support fine-grained server expansion very well. Today’s power-constrained
data center typically over-subscribes pieces of its power systems such as the PDU,
thereby creating power delivery bottlenecks. Adding renewable power capacity at data
center-level does not guarantee increased power budget for the newly added server
racks since the associated PDU probably has already reached its capacity limit.
Oasis does not synchronize renewable power to utility power due to three
considerations. First, at the PDU level, renewable power synchronization induces
voltage transients, frequency distortions, and harmonics. Data center servers are
susceptible to the power anomalies. As the impact of power quality issue on server
efficiency and reliability is still an open question, massively adding grid-tied renewable
power at rack level is debatable. Second, even if one chooses to synchronize
renewable power at data center facility-level, grid-tied renew-able power system can
cause efficiency problems. The newly added renewable power incurs many levels of
80
redundant power conversion before reaching server racks, resulting over 10% energy
loss [60]. Third, as required by UL 1741 and IEEE 1547, all grid-tied inverters must
disconnect from the grid if they detect power islanding. That is, these renewable power
systems must shut down if the grid power no longer present. As reported in a recent
survey, the U.S. data center experiences 3.5 times of utility power loss per year with an
average duration over 1.5 hours [60]. Thus, with synchronization based power
provisioning, data centers may lose power supply even when they need it most.
The intention of Oasis project is to facilitate server scale-out in power-/carbon-
constrained data centers. Oasis aims to become a non-intrusive, modularized
renewable power integration scheme which allows data center operators to increase
power capacity on-demand and reduce data center carbon footprint gradually.
6.2 Implementation of Oasis
In this section we describe our implementation of Oasis starting from scratch
(Figure 6-4A). Specifically, we focus on Oasis nodes, the key element of Oasis design.
As shown in Figure 6-4B, its structure consists of a power control hub, a server rack,
and a power management agent that coordinates them.
6.2.1 Power Control Hub
The power control hub (PCH) is the major hardware addition of Oasis nodes. It
integrates battery charger, power inverter, power supply switch panel, programmable
logic controller (PLC), and a human-machine interface (HMI) that allows easy system
diagnosis. The PCH is designed to manage multiple energy sources (i.e., renewable
power, utility power, and distributed battery devices). On its backside we provide three
electrical sockets that allows easy connection to utility power (AC), battery systems
(DC), and solar panels (DC).
81
A B Figure 6-4. Prototype of Oasis node. A) System prototype. B) Node architecture.
Our prototype system is powered by an array of solar module. Each module is a
270 Watt Polycrystalline panel from Grape Solar. The panel output is a complex
function of the solar irradiation, ambient temperature, and the load connected to it. To
harvest the solar energy, we use a maximal power point tracker (MPPT) in our power
control hub. The MPPT samples the output of the solar cells and applies proper control
to obtain the optimal solar energy generation. However, it is not an essential part of the
PCH. The MPPT is often a build-in module of the solar micro-inverter.
Oasis leverages distributed battery devices to provide temporary solar energy
storage. Energy storage devices, such as uninterruptable power supplies (UPS), are
widely used in data centers to address the risk of interruptions to the main grid. We
have customized a stack of nine 2Ah sealed lead-acid batteries for the power control
hub (scalable in the future). Our battery chassis can provide 5~10 minutes of backup
time, depending on the server load.
82
We convert the DC solar power to AC to match the output level of utility power.
The AC solar power and AC utility power are merged (but isolated) at a switch panel in
the PCH. We use high volt-age Omron relay to perform power switch and a Mitsubishi
FX2N programmable logic controller (PLC) is used to manage the switch behavior.
Finally, the PCH routes power to rack-level power distribution strip which further feeds a
cluster of server nodes. The entire rack is either powered by the utility power or
renewable power, depending on the status of the internal power switch.
We measured several key technical data of our current Oasis prototype. The
maximum charging current of our system is 4A, which is limited by the battery charger
for battery lifetime considerations. The PCH itself consumes static power (about 9
Watts) due to the HMI and PLC operations. The dynamic power loss is due to the heat
dissipation of power conversion systems.
6.2.2 Bridging Server and Power Supply
We enrich Oasis with sensing components that keep monitoring real-time solar
power output voltage and battery terminal voltage. These hardware agents could inform
us the renewable energy re-source conditions and the health status of batteries. Our
systematic checkup of power supply behaviors also offers a real-time profile of the
system’s energy utilization. This feature makes it possible to pin-point areas of high
energy usage in server farms. To facilitate real-time configuration and diagnosis, we
have designed human-machine interface (HMI) for each power control hub. As shown in
Figure 6-4A, it is a touch screen panel with built-in microprocessors that could display
graphics, animation, and interchange data to and from the PLC. Besides, the HMI
device also serves as an important portal for communication between the PLC and
power management agent.
83
HMI
PLC
(ModBus TCP/Server)(ModBus TCP/Client)
EthernetRS-232/485
Server OS
Workload Workload
Power Management Agent
Workload
Server System (The Load) Power Control Hub
Figure 6-5. Power management agent.
A key feature of the PCH is that it sets up the communication gateway between
power supply layer and server workload layer. This design is partly enlightened by the
energy management strategy in future smart grid, which focuses on intelligent control,
communication, and coordination across electrical loads, power electric interfaces, and
generators. Figure 6-5 illustrates the data communication scheme of Oasis. The power
management agent is a middleware lies between operating system and workloads.
We use Modbus protocol [61] for communication between the power
management agent and the power control hub. It is a widely used serial communication
protocol for industrial electronic devices due to its robustness and simplicity. There are
several other benefits of using Modbus. For example, the Modbus allows us to easily
scale up our deployment of nodes and share information among them. Besides, there is
no need to worry about the transmission failure of the Mod-bus instruction as the lower
layer TCP/IP protocol has provided the redundant checksum. Typical Modbus TCP
communication includes a Modbus server and a Modbus client. In this study, the power
management agent is the Modbus client (master) which initiates the communication
requests periodically through socket to the Modbus server (slave), i.e., the HMI.
6.2.3 Dynamic Energy Source Switching
Oasis nodes are able to dynamically switch between green power supply and
utility power. The PCH offers two power switching modes.
84
Autonomous Mode. This is the default mode. In this mode, Oasis nodes intend
to run autonomously, i.e., switching the load between solar power and utility grid based
on the given solar and utility power budget. The PLC in our power control hub defines
two atomic modules: SupplySense and SupplySwitch, as shown in Figure 6-6. While the
former module focusing on setting parameters that are used for making energy source
management decision, the latter one executes energy source switch.
Coordinated Mode. Oasis also provides servers the option of establishing their
own power supply switching policies. Our system currently allows two user-defined
operations: Utility Power Enforcement and Solar Power Enforcement. Users could
specify their preferred energy source at runtime, by calling the power management
agent. However, the execution of a power supply switching signal depends on battery
status, solar power output, and utility power availability. The user’s switching request
will be ignored if it violates the power budget or causes safety issues.
For every switch operation, the controller will first check the output of power
supplies to ensure that they work normally. The PCH configuration allows an overlap
between solar power and utility power when performing power supply switching. We
disconnect one energy source only when the other has been successfully working for 5
seconds. This mechanism helps to avoid potential switching failure that interrupts server
operation. Our controller also maintains appropriate voltage thresholds to prolong the
lifetime of backup power system. Figure 6-7 shows the measured battery voltage during
power switching. The battery starts to discharge at 12.5V (charging threshold) and stop
discharging when its voltage drops to 11.8V (discharging threshold). These thresholds
prevent batteries from entering deep-discharging or over-charging mode.
85
A
B
Figure 6-6. The control flow of power supply switching. A) The SupplySense module. B) The SuppleSwitch module.
Figure 6-7. Battery charging/discharging trace. The power control hub uses threshold
to avoid over charge/discharge a battery.
6.3 Optimized Oasis Operation (O3)
The flexibility of Oasis power provisioning architecture provides data center
owners plenty of room for improving the efficiency of their servers. In this section we
introduce Ozone, a power management scheme for Optimized Oasis Operation (O3).
Ozone enables Oasis users to take one step further towards the aim of efficient scale-
out data centers. Ozone seeks a balance between power supply management and
server load management. In addition to the basic function of Oasis, Ozone features a
new set of policies that intend to maximally take advantage of multiple energy sources
86
without heavily relying on any single energy sources. In addition, Ozone adaptively
adjusts server load based on power supply status to find the best design tradeoffs.
While many factors may affect Oasis operation, Ozone picks up six
representative factors for making control decisions. Three of them are power-related:
Utility Budget, Solar Budget, and Load Demand; two of them are battery-related:
Discharge Budget and Remaining Capacity; and one Switch parameter that specifies
which energy source is in use as primary power.
6.3.1 Managing Battery Lifetime
Batteries have a lifespan (typically 5~10 years) estimated by their manufactures.
These energy storage devices become no longer suitable for mission-critical data
centers after their designed service life. Aggressively discharging batteries will
significantly degrade their lifetime. However, even if batteries are not frequently used,
they still suffer aging and self-discharging problems.
Ozone uses an Ah-Throughput Model [62] for battery lifespan optimization. This
model assumes that there is a fixed amount of aggregated energy (overall throughput)
that can be cycled through a battery before it requires replacement. The model provides
a reasonable estimation of the battery lifetime and has been used in software developed
by the National Renewable Energy Lab [55].
During runtime, Ozone monitors battery discharging events and calculates
battery throughput (in Ah) based on the Peuket’s equation [63]: 𝐷𝑖𝑠𝑐ℎ𝑎𝑟𝑔𝑒 = 𝑇 ∙ 𝐼𝑎𝑐𝑡𝑢𝑎𝑙 ∙
(𝐼𝑎𝑐𝑡𝑢𝑎𝑙 𝐼𝑛𝑜𝑚𝑖𝑛𝑎𝑙⁄ )𝑝𝑐−1, where 𝐼𝑎𝑐𝑡𝑢𝑎𝑙 is the observed discharging current; 𝐼𝑛𝑜𝑚𝑖𝑛𝑎𝑙 is the
nominal discharging current, 𝑝𝑐 is the Peukets coefficient, and 𝑇 is the duration. Over
time, the aggregated battery throughput is given by 𝐷𝑎𝑔𝑔𝑟 = ∑ 𝐷𝑖𝑠𝑐ℎ𝑎𝑟𝑔𝑒𝑖𝑖 . To avoid
87
battery over-utilization, Ozone carefully sets a soft limit on battery usage at any given
time. At the beginning of each control cycle, Oasis nodes receive a Discharge Budget
which specifies the maximum amount of stored energy uses that will not compromise
battery lifetime. Assuming the overall battery throughput is 𝐷𝑡𝑜𝑡𝑎𝑙, our system calculates
a Discharge Budget as 𝐷𝐵 = 𝐷𝑡𝑜𝑡𝑎𝑙 − 𝐷𝑎𝑔𝑔𝑟. When the Discharge Budget is inadequate,
Ozone will switch server to utility grid or decrease load to avoid over-use of batteries.
6.3.2 Managing Backup Capacity
Distributed battery systems not only provide Oasis basic support for managing
time-varying renewable power, but also serve as uninterruptible power supplies for
scale-out servers. Maintaining necessary backup capacity is critical. It has been shown
that UPS capacity exceeded is one of the top root causes of unplanned outages in data
centers [60]. The backup capacity is the primary factor that determines UPS autonomy
(a.k.a. backup time). It is a measure of the time for which the UPS system will support
the critical load during power failure. In addition to the Discharge Budget, Ozone also
sets a limit on the minimum remaining capacity of batteries. Our system only uses 40%
of the installed capacity for managing renewable power shortfall (referred to as Flexible
Capacity). The remaining 60% battery capacity (i.e., Reserved Capacity) is only used
for emergency handling purpose. When the battery capacity drops below 60%, Ozone
will switch the server from renewable power supply to utility power supply.
6.3.3 Managing Server Performance
When renewable power output drops, there are always two power management
options available: (1) decrease server performance level to lower server power demand,
or (2) continue to operate at full speed and use energy storage to compensate the
power shortfall. To find the best design tradeoff between server performance and
88
system reliability, Ozone cooperatively controls power supply switching and server
speed. The idea is to adaptively select one of the two power management options
based on the observed Discharge Budget and the amount of Flexible Capacity.
When both Discharge Budget and Flexible Capacity are adequate, Ozone gives
high priority to server performance (i.e., run workload at the highest frequency) with the
support of stored green energy. If the system runs out of Discharge Budget but still have
Flexible Capacity, Ozone allows the server to keep using green energy with reduced
processing frequency. If the stored energy drops below 60% of its installed capacity,
Ozone switches the load to the utility power side to avoid low battery backup.
6.4 Experimental Methodology
We develop an evaluation framework using our prototype system. The framework
is configured into three layers: the Power Budget Control Layer, the Oasis Operation
Layer, and the Data Collection and Analytic Layer.
In the Power Budget Control Layer, we feed the system with pre-defined power
budget. We use the peak server power demand as default utility power budget of our
system. To ensure fare comparison, we collect real-world solar power traces and use it
as renewable power budget for all the experiments.
Figure 6-8. The power control decision tree of Ozone. ‘Y’: switch the power supply;
‘N’: do not switch; ‘TBD’: server speeds depends on actual power budget.
89
Table 6-1. Computing platform setup for Oasis. Computing Nodes Front-End Server
CPU Intel Core i7-2720QM, 2.2G, 45W AMD Opteron 4256 EE, 2.5G, 32W RAM 8 GB, registered 16 GB, registered Disk Seagate Barracuda 7200RPM, 500 G Seagate Barracuda 7200RPM, 1000 G M/B SuperMicro ITX Socket G2 M/B SuperMicro ATX Socket C32 M/B Network N/A TP-Link TL-SF1024 24-Port 10/100M
In the Oasis Operation Layer, we setup four 1U rack-mounted servers as
computing load, as shown in Table 6-1. They are high-performance lower-power
computing nodes that use Intel Core i7-2720QM 4-core CPU as processing engine.
With the Intel Turbo Boost Technology, these processors support up to 3.3 GHz
operating frequency. We deploy Xen 4.1.2 with Linux kernel 2.6.32.40 on each server
node. Both para-virtualization and hardware virtualization are used to support different
virtual machines with various memory size. Multiple virtual machines are booted to
execute different workloads on each server. We enable the relocation feature of VM in
Xen and live migrate the virtual machine by using command (xm migrate DOMID IP -l).
Xen power management is also enabled to dynamically tune the vCPU frequency. Our
system is configured with the on-demand frequency scaling governor. We set the
minimum frequency as 0.8GHz and the normal frequency as 2.2 GHz.
We choose various data center workloads from Hibench [64] and CloudSuite
[65], Hibench consists of a set of representative Hadoop programs. CloudSuite is a
benchmark suite designed for emerging scale-out applications in datacenters. As shown
in Table 6-2, we select ten workloads from five roughly classified categories.
In the Data Collection and Analytic Layer, we deploy front-end network server to
communicate with the server cluster through a TP-Link 10/100M rack-mounted switch.
The network server uses an AMD low power 8-core CPU with 16GB installed memory.
We write system drivers using Linux socket to enable data communication between
90
front-end server and the Oasis node. We store the collected battery charging and
discharging statistics and the measured server power consumption data in a log file. We
assume the battery has a cycle life of 5000 times and a maximal service life of ten
years. We also use Watts UP Pro power meter [66] for our measurement in energy-
related experiments. This power meter is able to display instantaneous power
consumption with relatively high accuracy (1.5%). It also provides internal memory for
storing up to 120K history power data.
Table 6-2. Evaluated cloud workloads for Oasis.
Abbreviation Workload Category
Sort Sort program on Hadoop Micro Benchmarks WC Word count program on Hadoop Micro Benchmarks Rank Page rank algorithm of Mahout Web Search Nutch Apache Nutch indexing Web Search Bayes Bayesian classification Machine Learning KM K-means clustering Machine Learning Web Web serving Internet Service Media Cline-server media streaming Internet Service YCSB Yahoo! cloud serving benchmark Cloud Application Test Software testing Cloud Application
6.5 Evaluation Results
In this section we evaluate the impact of various power supply switching and
server adaptation schemes. To be more specific, we evaluate three kinds of Oasis
power management schemes: Oasis-B, Oasis-L, and Ozone. Oasis-B is a battery
oriented design which aggressively uses battery to store solar energy for powering
server clusters. The Oasis-L mainly focuses on the load power scaling capability of
servers. It first uses frequency scaling to decrease load power, then leverage battery to
compensate the power shortfall. In contrast to Oasis-B and Oasis-L, Ozone focuses on
balanced usage of server load adaptation and the stored renewable energy. In the
following sub-sections, we first investigate the performance of various control schemes.
91
We then discuss their impacts on battery lifetime and emergency handling capabilities.
Afterwards, we evaluate the energy usage profile of different schemes. Finally, we
discuss the cost issue of Oasis.
6.5.1 Load Performance and Energy Efficiency
Job turn-around time is a crucial metrics for emerging data-analytic workload in
scale-out data centers. Figure 6-9 shows the in-crease in job turn-around time under
both high-variability and low-variability solar power generation scenarios. In Figure 6-
9A, the mean turn-around time of Oasis-B, Oasis-L and Ozone is increased by 0.5%,
5.4%, 0.6%, respectively. In Figure 6-9B, the turn-around time of Oasis-B, Oasis-L and
Ozone is increased by 0.4%, 4.7%, 1%, respectively. As can be seen, Oasis-B shows
the best performance. This is because Oasis-B trades off battery lifetime for server
performance. In contrast, Oasis-L shows much higher performance degradation as it
frequently lowers the processing frequency to match the inadequate renewable power
budget. In Figures 6-9A and 6-9B, the results of Ozone are very close to Oasis-B. On
average, Ozone yields less than 1% performance degradation compared to Oasis-B,
which heavily uses battery to provide power shortfall.
Data centers with heavy data-processing computing task often consume large
amount of energy. Leveraging green energy to provide additional power could save
utility power bills considerably and lower the negative environmental impact of carbon-
constrained data centers. In Figure 6-10 we evaluate data center green energy
utilization as the ratio of renewable energy usage to overall IT energy consumption.
While Ozone yields impressive system performance, battery lifetime, and battery
backup capacity, it shows relatively lower green energy utilization. Compared to Oasis-
B, Ozone yields 11% less renewable power rate when renewable power varies
92
significantly and 8% less renewable power rate when renewable power generation is
high. The reason Oasis-B shows high green energy usage is that it heavily uses battery
to harvest renewable energy. In contrast, Oasis-L aggressively scales load power
demand to match the renew-able power generation.
A B
Figure 6-9. The execution latency due to server performance scaling. A) High renewable power variability. B) Low renewable power variability.
A B
Figure 6-10. The ratio of green energy usage to overall power consumption. A) High renewable power variability. B) Low renewable power variability.
6.5.2 Battery Service Life and Backup Capacity
In Figure 6-11 we estimate the battery service life based on detailed battery
profiling information. When renewable power supply varies significantly, the operation of
server nodes typically requires substantial support from the energy storage. As a result,
the predicted lifetime is much shorter than the designed service life, as shown in Figure
6-11A. On average, the battery lifetime of Oasis-B, Oasis-L, and Ozone is 3.9 years, 6.2
years, and 6.0 years, respectively. Due to the over-use of battery systems, the battery
lifetime of Oasis-B is only 63% of Oasis-L. In contrast, the battery lifetime of Ozone is
97%. Compared to Oasis-B, Ozone increase the battery lifetime by more than 50%.
93
When renewable power supply becomes adequate, as shown in Figure 6-11B,
the battery service of all the three power management schemes increases significantly.
This happens because batteries are not discharged for most of time. However,
commercial batteries cannot last for 20 years even if they are under-utilized. Many other
issues such as aging and leakage will become the dominant factors for lifetime
estimation if the batteries are used for an extended duration. In fact, batteries are used
more frequently since renewable power system does not maintain peak output.
A B
Figure 6-11. The estimated battery lifetime calculated from battery usage profile. A) High renewable power variability. B) Low renewable power variability.
Figure 6-12 illustrates the problem of uneven battery usage. Note that when
renewable output is constantly high, the distributed battery system is rarely used. In
contrast, when renewable power fluctuates severely, batteries start to show heavy
charging and discharging activities. Therefore, one can actually save plenty of
Discharging Budget in scenarios like Figure 6-12B. This amount of Discharging Budget
saving can be further used in scenarios like Figure 6-12A to provide necessary load
power support. In fact, the Oasis user can opportunistically leverage stored energy
(even if current Discharging Budget is zero) to boost system performance without
significantly affecting the battery service life.
94
A
B
Figure 6-12. Battery discharging profile and power demand traces. The workload used is Nutch page indexing. A) High renewable power variability. B) Low renewable power variability.
When data center servers scale out, their backup up power should be increased
accordingly. While instantaneous workload performance boost is important, maintaining
necessary backup capacity is even more important. A low battery backup capacity can
pose significant risk since the backup generator may not be ready to pick up the load. In
Figure 6-13 we show the average UPS capacity during runtime for different power
management mechanisms. Ozone could maintain around 73% backup capacity when
renewable power variability is high and about 98% backup capacity when renewable
power variability is low. Oasis suffers increased numbers of charging and discharging
cycles in circumstances that renewable power varies significantly. Therefore, the
backup capacity is low for all the three power management schemes in Figure 6-13A.
Without setting a limit on the battery usage and the minimum battery stored capacity,
the battery backup time can drops by 75% (i.e., Oasis-B).
95
A B
Figure 6-13. The observed average battery backup capacity in our UPS systems. A) High renewable power variability. B) Low renewable power variability.
A B Figure 6-14. A cost breakdown of Oasis. A) Summary of our prototype. B) Estimation
for a 5KW standard server rack.
6.6 Cost Analysis of Oasis
This section discusses the system cost of Oasis. Figure 6-14 presents two pie
charts that show the cost breakdown. In Figure 6-14A we evaluate our Oasis node
prototype. As can be seen, solar panel is the most expensive component in our system,
followed by the PLC module, the power inverter, and battery. In Figure 6-14B we
estimate the cost of Oasis design for a 40U standard server with moderate density
(<10KW). We assume the rack is populated with 20 SuperMicro 1017R Xeon 1U
servers. While PLC and HMI are both major cost components in our prototype system,
they do not need to scale up when more servers are added. In contrast, solar panels,
batteries, and inverters need to increase their capacities.
96
Centralized renewable power integration has relatively low initial cost due to the
scale effect. Recent report estimates that small-scale PV panel (around 5KW, for
residential use) has an installed price of $5.9/W, while large-scale PV panel (several
hundred KW) has a lower price of around $4.74/W [68]. In addition, solar power inverter
accounts for about 10%~20% of the initial system cost [69]. Central inverters in several
MWs level are often cheaper compared to micro inverters (typically < 10KW) used in
Oasis. The former costs around $0.18/W, while the later costs around $0.5/W [69].
Oasis allows users to gradually increase the installed renewable power capacity,
thereby eliminating over-provisioning. Oasis users can also take advantage of the ever-
decreasing component cost to further lower costs. It has been shown that the installed
prices of U.S. residential and commercial PV systems declined 5%~7% per year, on
average, from 1998~2011, and by 11%~14% from 2010~2011, depending on system
size [69]. The cost of micro-inverter also de-creases by 5% yearly [69].
Figure 6-15 illustrates how Oasis design helps to improve the overall cost-
effectiveness of renewable energy powered scale-out data centers. We assume that
Oasis users evenly increase their deployment of Oasis node with a ten-year scale-out
plan. (e.g., equip 10% of the data center servers with solar power system every year).
For the cost of solar power system, we use a conservative decline rate of 6% per year,
and an optimistic decline rate of 12% per year. We calculate electricity cost savings of
renewable energy powered systems based on real historical solar power traces. We use
hourly solar irradiance measurement data (January 2003 ~ December 2012, 24 hours a
day) provided by the NREL Solar Radiation Research Laboratory [70]. We assume the
utility power is $0.1/kWh and data centers sell excess renewable power to the utility.
97
Figure 6-15. Cost trends of deploying Oasis for scaling out.
Figure 6-15 shows the cost overhead (total additional cost due to renewable
power integration) for different design scenarios. The results are normalized to the one-
time expenditure of data center-level renewable power integration. For conventional
green data center that uses centralized renewable power integration, one can expect to
get 17% investment return (due to the electricity cost savings) after ten years. However,
this estimation is optimistic as the utility grid typically uses negotiated renewable power
feed-in tariff that has a lower purchase price. In addition, for safety reasons, there is
also a limit on the maximum amount of renewable power that can be synchronized with
the utility power. The overall cost of Oasis is close to the conventional centralized
design if solar power cost decreases by a conservative rate of 6% per year. If solar
power cost declines faster, namely, 12% per year, Oasis could result in 25% less
overhead cost, compared to a centralized design.
Note that there is no one-for-all universal design. Green data-centers with
centralized solar power system may be a better choice if one has a firm goal of
renewable power integration capacity and a confidence in future load power demand.
However, if one wants to enhance data center operational resiliency and seeks to
gradually incorporate renewable energy to avoid over-committing capital, Oasis
provides an attractive alternative.
98
CHAPTER 7 SUSTAINABLE COMPUTING IN THE SMART GRID AND BIG DATA ERA
Looking ahead, the attributes of workload application and energy source will
undoubtedly influence the way we compute. In the future when computer architecture
meets big data workload and smart grid techniques, it requires rethinking of
conventional computing system design approaches.
First, as we move toward a smarter grid, data centers are expected to be
powered by hybrid renewable energy systems that combine multiple power generation
mechanisms [71]. For example, it may include intermittent power supplies such as wind
turbines and solar panels, controllable power supplies such as fuel cells and gas
turbines, and various backup power system such as batteries and super-capacitors.
With an integrated mix of complementary power provisioning methods, one can
overcome the limitations of each single type of power supply, thereby achieving better
energy reliability and efficiency. However, a common limitation of prior proposals is that
they mainly focus on certain specific type of green power supplies. As a result, they can
hardly gain the maximum benefits from hybrid renewable energy systems, and
consequently yield sub-optimal design tradeoffs.
Second, although much of the computing resources today are hosted in data
centers, a tremendous amount of datasets are generated from distributed machines,
monitors, instruments, and various types of sensor networks. For example, a large
network that consists of ten thousands tiny nodes can easily generate over 5 terabytes
(TB) of data within a week [72]. In addition, today’s fast-growing scientific datasets (e.g.,
climate data and genome data) are typically distributed among many stations and
research institutions around the world. Such wide-area collaboration on location-
99
dependent datasets normally requires a routine data sharing of tens of petabytes (PB)
every year [73]. Consequently, the challenges of analyzing these geographically
dispersed data sets are increasing due to significant data storage overhead, time-
consuming data aggregation and the escalating energy needs. Rather than constantly
move a tremendous amount of raw data to remote warehouse-scale systems, it would
be beneficial to bring computation to where data is located.
This chapter contains two case studies that explore the opportunities to benefit
from various green computing systems in the incoming smart grid and big data era. We
first propose GreenWorks, a power management framework for green high-
performance computing data centers powered by renewable energy mix. Afterwards, we
present InSURE, an intelligent design of stand-alone green server clusters for locally
processing distributed big datasets generated in the field.
7.1 Green Data Centers Powered by Hybrid Renewable Energy Systems
In the smart grid era, data centers must increase their awareness of the
attributes of power supplies to achieve the best design trade-offs. There are three types
of renewable power supplies that we can leverage to power a data center: Some green
power supplies, such as solar panels and wind turbines, are affected by the availability
of ambient natural resources. They are referred to as intermittent power supply. Several
emerging power supplies – including fuel cells, bio-fuel based generators – can offer
controllable output by burning various green fuels. We refer to them as baseload power
supply since they can be used to provide stable renewable power to meet the basic
data center power demand (e.g., idle power). In addition, energy storage devices such
as batteries and super-capacitors are also critical components that provide backup
power supply. They can be used to temporarily store green energy.
100
Figure 7-1. High-level system architecture for GreenWorks.
Data centers in the smart grid era are expected to be powered by hybrid
renewable energy systems that combine all the three types of power supplies. They are
typically implemented as small, modular electric generators (called micro-sources) near
the point of use. This section explores diversified multi-source power provisioning for
green high-performance data-centers. The intention of this work is to provide an initial
power management framework for data centers powered by renewable energy mix.
7.1.1 Hierarchical Power Management Framework
GreenWorks is a hierarchical power management scheme that is tailored to the
specific timing and utilization requirements of different energy sources. It provides
coordinated power management across intermittent renewable power supplies,
controllable base-load generators, onsite batteries, and data center servers.
As shown in Figure 7-1, GreenWorks comprises two key elements: the green
workers and the green manager. The former are platform-specific power management
modules for managing different types of micro-sources and the later coordinates these
modules. In this study we define three types of green workers: baseload laborer (B),
energy keeper (E), and the load broker (L). We adopt typical micro-grid power
distribution scheme for managing various renewable energy resources.
101
Figure 7-2. The GreenWorks power management hierarchy.
Although the hybrid renewable energy systems are often centrally installed at the
data center facility level, improving the overall efficiency requires a multi-level,
cooperative power management strategy, as shown in Figure 7-2.
Baseload Laborer. In the top tier of the hierarchy is the baseload laborer. It
controls the output of controllable generators. We put the load laborer at the data center
facility level since it is where the baseload generator is integrated.
Server Load Broker. The load broker manages the intermittent renewable
power supply at the cluster/PDU level. At this level, dynamic voltage and frequency
scaling (DVFS) shows impressive peak power management capabilities [43] and could
be leveraged to manage the supply-load power mismatch.
Battery Energy Keeper. The energy keeper resides in the third tier of the
hierarchy. This allows us to provide backup power directly to server racks if local
demand surge happens or power budget drops. It also monitors the capacity utilization
and the health status of the battery packs for optimization purpose.
7.1.2 Multi-Source Driven Power Management
The main advantage of our multi-level power management framework is that it
facilitates cross-source power optimization. GreenWorks features a novel three-stage
coordination scheme to balance the usage of different types of energy sources [74].
102
Stage I: Adequate Power Supply Budget. When allocating the additional
renewable power budget across server nodes, the green manager will always give
priority to jobs that have higher job execution time increase (ETI). Specifically, our green
manager uses a job acceleration scheme which opportunistically boosts the processing
speed/frequency (i.e., over-clocking) to take advantage the additional renewable power
budget [74]. It allows a processor to enter a performance state higher than the specified
frequency if there is enough thermal/power headroom and if it is enabled by the system.
Stage II: Moderate Power Supply Drop. Our system enters Stage-II when it
senses inadequate power supply. Unlike prior designs which heavily rely on either load
shedding or backup power, we use a balanced power management. GreenWorks
choose to decrease server processing speed or request stored energy from the energy
keeper, depending on whichever yields the best design tradeoffs [74].
Stage III: Significant Power Supply Drop. The Stage-III is an emergency state
since in this scenario the green manager might put the load into minimum power state
or use reserved UPS capacity to avoid server shutdown. GreenWorks uses a deadline-
aware load shedding to achieve a better tradeoff between UPS energy and job
execution time increase (ETI). The green manager first checks the current ETI values of
all the jobs for load shedding opportunities. It calculates a Time Budget which evaluates
if a job could meet its deadline in the future with frequency boosting techniques [74]. If
the given job has enough Time Budget, our controller will incrementally reduce its CPU
frequency until it reaches its lowest speed. It will put server nodes into low power states
in a round-robin fashion if the demand-supply discrepancy still exists. Finally, we
release the reserved UPS energy if necessary.
103
Figure 7-3. Details of our three-layer simulation platform.
7.1.3 Verification Platform and Results
Our simulation is configured into three layers for modeling the entire system from
the job dispatching behavior to the renewable energy mixes. It uses discrete-event
simulation to process a chronological sequence of job submissions. It also simulates the
power system on per-second time scale which is in tune with our workload scheduler.
We compare GreenWorks to two baselines: Shedding and Boosting. Shedding
always defers workload when renewable energy drops [12] [21]; Boosting emphasis the
role of energy storage devices [15] [16]. Both baselines adjust baseload output at each
end of the control period. The only difference between the two is that Shedding gives
priority to load scaling, while Boosting gives priority to UPS stored energy.
Execution Time. Figure 7-4 shows the average job execution time increase. On
average, the job execution time increase of Shedding, Boosting and GreenWorks are
5.4%, 2.1%, and 2.4%, respectively. Compared to Shedding, Boosting shows less
execution time increase since it trades off UPS capacity for performance. Figure 7-5
shows the maximum increase of job turnaround time which is calculated as the average
execution time increase of the 5% worst cases. The worst-case result of Shedding is
28%. Because GreenWorks apply a performance limit when shedding load, it reduces
the maximum job execution time increase by 33%, compared to Boosting (18%).
104
A B
Figure 7-4. The impact of GreenWorks on job execution time. A) Average ETI across all the job requests. B) Maximum ETI for the worst 5% jobs.
Battery Lifetime. In Figure 7-5, GreenWorks shows a near-threshold battery life
(8.3 years). It means our power management can maximally leverage batteries without
degrading their life significantly. In contrast, Boosting shows a mean lifetime of 6.7
years; and Shedding shows a mean lifetime of 19.7 years. Typically, the battery lifetime
is not likely to exceed 10 years [75]. The reason Shedding over-estimates battery life is
that the system underutilizes batteries. Since batteries may fail due to aging problems
and self-discharging issues, it is better to fully utilize it.
UPS Backup Time. GreenWorks can further optimize the mean UPS autonomy
time (a.k.a. backup time). It is a measure of the time for which the UPS system will
support the critical load during an unexpected power failure. Figure 7-6 shows the mean
normalized UPS autonomy time throughout the operation duration. On average, the
mean autonomy time is: Shedding (88%), Boosting (70%), and GreenWorks (78%). In
Figure 7-7 we plot the cumulative distribution function (CDF) for the normalized UPS
autonomy time. The CDF curve of GreenWorks lays nicely between our two baselines.
GreenWorks could ensure rated backup time (the discharge time of a fully charged
UPS) for 20% of the time. Shedding maintains its rated backup time for 50% of the time,
while the number for Boosting is only 10% due to aggressive battery usage.
105
Figure 7-5. Battery lifetime estimation
based on its usage log. Figure 7-6. Average UPS backup time
(normalized to the rated value).
Figure 7-7. Cumulative distribution function (CDF) for UPS autonomy time.
7.2 Towards Sustainable Power Provisioning for In-situ Server Systems
Although much of the computing resources today are hosted in data centers, a
tremendous amount of datasets are generated from distributed machines, monitors,
instruments, and various types of sensor networks.
Rather than constantly move a huge amount of data to a central data warehouse
for aggregation and processing, we instead explore a fundamentally different approach:
tapping into in-situ server systems. The idea is to bring servers to where data is located
to pre-process part, if not all, of the datasets. For instance, these servers can be used to
eliminate duplicate copies, compress logs, normalize data formats or classify data
types. The concept of in-situ data processing has been used in the high-performance
computing (HPC) community to solve the I/O overhead for compute-/memory- intensive
workloads [76] [77]. In this work, we repurpose this concept to design onsite server
systems that can accelerate or facilitate pre-processing of distributed big datasets.
106
Our interest in in-situ server systems also arises out of the fact that modern data
centers are power-constrained, particularly when they employ power over-subscription
techniques to reduce cost. In the past five years, 70% companies have to build new
data centers or significantly renovate existing facilities to handle the ever-growing traffic
[56]. Meanwhile, recent studies are forced to aggressively discharge backup batteries to
provision more server under existing power budget [17] [78]. As data continues to flood
into data centers, it is not unusual that the utility power feeds are at full capacity and
data centers do not have enough power to accommodate their growth [79] [80].
There are also significant challenges associated with power provisioning for in-
situ servers. Different from sensor nodes, computer servers require much higher power
budget and cannot solely rely on batteries. In remote areas or rural locations that are
environmentally sensitive, the construction and operation of transmission lines could be
prohibitive [81]. Even if utility power access is not an issue, conventional server power
provisioning methods often violate local environmental quality regulations [82].
We present InSURE: in-situ server systems using renewable energy. Such in-situ
server system is well complementary to cloud computing in the incoming big data era. It
can greatly reduce data processing overhead in public cloud and private network, with
zero carbon emission. As Figure 7-8 shows, we explore the opportunity to benefit from
data pre-processing using a group of inexpensive servers that are placed near the data.
Specifically, we focus on standalone systems powered by local renewable energy
sources. It has been shown that about 85% data processing tasks can be deferred by a
day; 40% task can be deferred by a month [83]. Therefore, even if the renewable power
is intermittent and time-varying, we can still leverage it for data pre-processing.
107
Figure 7-8. In-situ server system as an ancillary to future cloud.
7.2.1 Standalone Green Server Clusters
Since batteries are widely used as energy buffer that links renewable power
supply with load [27], their efficient use is the crux of high data processing throughput.
Without appropriate control, battery charging process (using green energy) could incur
unnecessary slowdown and, consequently, cause data processing delay. Besides, peak
load could cause significant battery voltage drop even if there is still enough renewable
energy stored. This can disrupt server operation and greatly degrade productivity. To
this end, we have designed a cross-layer power management scheme that is tailored to
the specific behavior of standalone in-situ systems. It highlights two techniques:
Reconfigurable distributed energy storage: We synergistically integrates
existing power switch network with emerging distributed energy storage architecture. It
allows us to dynamically adjust the size of energy buffer for better design trade-offs.
Joint spatial-temporal power management: Our spatial control could
intelligently balance the longevity of battery systems and ensure the optimal charging
effectiveness under varying renewable power generation conditions. Our temporal
control scheme cooperatively manages server power demand and battery discharge
current to improve battery discharge efficiency and reduce load shedding frequency.
108
Figure 7-9. In-situ server cluster with reconfigurable battery array. Although we focus on stand-alone system, our design also supports a secondary power (e.g., local micro-grid, if available).
Our reconfigurable energy buffer is shown in Figure 7-9, three power switches
(P1, P2, and P3) are used for managing the battery cabinets to provide different voltage
outputs and ampere-hour ratings to the load. For example, if P1 and P3 are closed
while P2 is open, batteries are connected in parallel. If switches P1 and P3 are open
while P2 is closed, batteries are connected in serial. A set of sensors S1 is used to
monitor the runtime operation condition of each battery cabinet. All switches and
sensors are managed by a PLC controller.
The energy buffer of InSURE supports a variety of operating modes, as shown in
Figure 7-10. Based on the states of energy storage, we categorize them into four types:
Offline, Charging, Standby, and Discharging. In the Offline mode, batteries are
disconnected from the server load for system protection purposes. In the Charging
mode, onsite renewable power, if available, is used for charging batteries with the best
achievable efficiency. Our system put batteries online when they are charged to a pre-
determined capacity (i.e., 90%). In the Standby and Discharging modes, our system
mainly relies on renewable energy (either directly generated from the onsite green
generator or previously stored in batteries) to power green server clusters.
109
Figure 7-10. The energy flow scenarios for InSURE.
Figure 7-11. Operating modes transition of InSURE energy buffer.
The transition between various operating modes is shown in Figure 7-11. The
energy buffer is able to adjust its operating mode based on the stored renewable energy
budget, server power demand, and battery health conditions.
7.2.2 System Implementation and Verification
We have built a full-system prototype of InSURE as our verification platform. It is
a three-tier hierarchical system, as illustrated in Figure 7-12. It is composed of three
primary modules built from scratch: (1) a reconfigurable battery array, (2) a real-time
monitoring module, and (3) a supply-load power management/coordination node.
Reconfigurable Battery Array. Our customized battery system is composed of
six 12 V lead-acid batteries and a relay network. We use six 10A/24V DC relays as the
power switches for reconfiguring the battery array. Each battery is managed
independently using a pair of two relays (charging switch + discharging switch). These
relays provide satisfactory mechanical life (10M operations). We also use a Siemens
S7-200 CPU224 PLC module as the controller for our reconfigurable battery system. Its
digital output is connected to the relay network and can energize or de-energize the coil
of relay to perform battery switching.
110
Figure 7-12. The structure of our verification platform.
Real-Time System Monitoring. The monitoring module notifies our system
whenever the battery configuration profile changes. Each battery in the system is
equipped with a voltage transducer and a current transducer. Their outputs are further
sampled by two PLC extension modules. All the analog readings will be stored in
several specific registers in the PLC. We use an HMI panel to link battery system and
the coordination server node via Ethernet using the Modbus TCP protocol [61].
Power and Load Coordination. The top hierarchy of our design consists of a
power coordination module. This module is implemented on a separate server node.
We have designed a power supply switching API and a server control API. The former
API provides necessary communication interface that allows the system to select its
power source during runtime, if it is necessary. The latter API is used to manage load
power through frequency scaling, power state control, and virtual machine migration
7.2.3 Analyzing In-Situ Green Computing
To understand the power behavior of inSURE when handling bulk data using
self-generated solar energy. We investigate a typical system power behavior trace that
are collected from our prototype, as shown in Figure 7-13. We have marked five typical
regions (Region-A ~ Region-D) that frequently appears in our everyday operation.
111
Figure 7-13. Solar power budget trace and battery voltage trace Top figure shows a
full-day operation and the bottom provides a zoomed view.
Our standalone in-situ system starts to charge a selected subset of battery in the
morning, as shown in Region-A. During this period, the battery terminal voltage
gradually increases until it researches a preset value.
Figure 7-13 also shows a zoomed view of a fraction of the system trace. As we
can see, Region-B incurs a great deal of solar usage surges/peaks. This is because our
system uses a Perturb and Observe (P&O) maximum renewable power tracking
mechanism [84]. In Region-B, each green peak is a result of our peak power tracking
control. It will not cause overloading since the battery can temporarily provide the power
short fall. Region-C shows our temporal control. In this region the solar power demand
of our in-situ system is significantly larger than the maximum solar budget, and battery
is in discharging mode. If the discharge current is too high, our system will trigger power
capping. As a result, the power demand drops at the end of this region.
In the Region-D of Figure 7-13, we are in the most desired region. Renewable
power is adequate and our in-situ server system can get the most benefit from
renewable energy powered data-preprocessing. In contrast, Region-E is an unfavorable
regions as severely frequenting solar power budget can cause many load power peaks.
We will adjust our peak power tracking mechanism to address this issue.
checkpoint and suspend servers
initial battery charging
supply-demand matching
severely fluctuating solar power
6:54 AM 7:59 PM
9:28 AM 11:18 AM
A
D
E
Cusage
usage
Bpower tracking
b
b v
v
112
Table 7-1. Key variables collected from the inSURE system log. We show three typical renewable power generation scenarios.
kwh Usage Total(Useful)
Power Ctrl. Freq.
Server on/off Cycles
VM Ctrl. Freq.
Battery Volt. σ
Sunny & Aggr. Ctrl. 6.7 (6.4) 12 8 8 1.05 Sunny & Non-Aggr.* 6.5 (5.9) 47 16 42 0.93 Cloudy & Aggr. Ctrl. 5.5 (5.2) 10 8 15 1.03 Cloudy & Non-Aggr. 5.0(4.2) 51 20 51 0.92 Rainy & Aggr. Ctrl. 2.8(2.5) 10 8 11 1.04 Rainy & Non-Aggr. 2.6(2.1) 33 15 38 0.93
*A non-aggressive battery management has a limit on battery discharge throughput.
We further investigate the system behaviors of InSURE by looking at the detailed
system logs. As we can see from Table 7-1, the effective energy consumption usually
does not equal to the overall load energy consumption. This is because VM check-
pointing operations and the on/off cycles of server systems also consume energy. On
average the effective energy utilization is about 89% of the total load energy
consumption. The non-aggressive server power management yields about 86%
effective energy consumption, and about 94% total load energy consumption, both
compared to the aggressive schemes. This is because the nonaggressive server power
management trades off energy utilization for battery life-time.
In Figure 7-14 we further evaluate the overall impact of our spatio-temporal
power management on in-situ server systems. We compare our system with
conventional design that cyclically charge and discharge battery (i.e., aggressive battery
management). As shown in the figure, our spatio-temporal control has bigger impact on
datacenter availability and battery capacity (i.e., SOC). The improvement over a con-
design mainly comes from two sources. First, the temporal control prevents battery from
sudden, fast capacity drop (voltage drop). Second, by dynamically concentrates the
limited renewable power generation on a smaller set of batteries to achieve fast battery
charging, our spatial control can greatly accelerate the charging speed of batteries.
113
A B C
Figure 7-14. Expected optimization results from our spatio-temporal management. A) Service availability. B) Battery state of charge. C) Battery lifetime.
Another observation is that the power budget level has a significant influence on
the optimization effectiveness of our technique. Our system shows about 41% service
availability improvement compared to conventional design under low solar generation,
while the improvement can reach 51% under high solar generation. If we look at each
individual workload, we can find that our system always shows larger availability
improvement under low solar generation level. This is mainly because the benefit of our
spatial power management often increases when renewable power is inadequate.
We also observe a battery lifetime improvement of approximately 21~24% across
our workload mix. The optimization of battery service life is mainly a result from battery
discharge capping and battery discharge balancing across the battery cabinet. The
significant lifetime improvement also indicates that with our design, a larger amount of
solar energy are directly used by server load (i.e., bypasses the battery), thereby
reducing the electrical ware-out on batteries.
114
CHAPTER 8 CONCLUSIONS
This dissertation envisions the long-term competitiveness of designing green
energy aware computing systems. We demonstrate that a cross-layer and cross-
component power management scheme that span workload applications, computer
hardware, and energy sources is the crux of achieving high efficiency and high
performance in such design. The proposed techniques can greatly contribute to
enabling computing systems to stay on track with their historic performance scaling
trend, and thus, paves the way for the ultimate solution to the energy and environmental
problems of modern IT infrastructures.
This work is only a glimpse into the possibilities opened by emerging computing
systems and energy systems. The expected growth in data center heterogeneity and
power provisioning diversity will unavoidably increase the design complexity of green
computing platforms. Future work on a holistic load/power management scheme within
and across green data centers will allow us to maximize the benefits we can get from
the synergy of computing facility and energy systems.
115
LIST OF REFERENCES
[1] C. Belady, (March 2011). Projecting Annual New Data center Construction Market Size, accessed on May 2014. Global Foundation Services, White Paper, [Online]. Available: http://www.business-sweden.se/PageFiles/9118/Projecting_Annual_New_Data_Center_Construction_Final.pdf
[2] J. Koomey, (August 2011). Growth in Data Center Electricity Use 2005 to 2010, accessed on May 2014. Analytics Press, White Paper, [Online]. Available: http://www.analyticspress.com/datacenters.html
[3] G. Boccaletti, M. Löffler, and J. Oppenheim, (October 2008), How IT can cut carbon emissions, accessed on May 2014. [Online]. Available: http://storage.globalcitizen.net/data/topic/knowledge/uploads/2011110295634705.pdf
[4] Greenpeace International, accessed on May 2014. [Online]. Available: http://www.greenpeace.org/usa/en/
[5] IEEE Guide for Array and Battery Sizing in Stand-Alone Photovoltaic (PV) Systems, IEEE Standard 1562™-2007.
[6] S. Kaxiras and M. Martonosi, Computer Architecture Techniques for Power-Efficiency, Morgan & Claypool, 2008, pp. 1-185.
[7] R. Teodorescu and J. Torrellas, “Variation-aware application scheduling and power management for chip multiprocessors,” in Proc. 35th Int. Symp. on Computer Architecture (ISCA’08), 2008, pp. 363-374.
[8] C. Isci, A. Buyuktosunoglu, C.-Y. Chen, P. Bose and M. Martonosi, “Analysis of efficient multi-core global power management policies: maximizing performance for a given power budget,” in Proc. 39th Int. Symp. on Microarchitecture (MICRO’06), 2006, pp. 347-358.
[9] J. Lee and N. Kim, “Optimizing throughput of power- and thermal-constrained multicore processors using DVFS and per-core power-gating,” in Proc. 46th Design Automation Conference (DAC’09), 2009, pp. 47-50.
[10] N. Sharma, S. Barker, D. Irwin, and P. Shenoy, “Blink: Managing server clusters on intermittent power,” in Proc. Int. Conf. on Architectural Support for Programming Languages and Operating Systems (ASPLOS’11), 2011, pp. 185-198
[11] N. Deng, C. Stewart, and J. Li. “Concentrating renewable energy in grid-tied data centers,” in Proc. Int. Symp. on Sustainable Systems and Technology (ISSST’11), 2011, pp. 1-6.
116
[12] I. Goiri, K. Le, T. Nguyen, J. Guitart, J. Torres, and R. Bianchini, “GreenHadoop: Leveraging green energy in data-processing frameworks,” in Proc. 7th European Conference on Computer Systems (EuroSys), 2012, pp. 57-70.
[13] Y. Zhang, Y. Wang, and X. Wang. “GreenWare: Greening cloud-scale data centers to maximize the use of renewable energy,” in Proc. 12th International Conference on Middleware (Middleware’11), 2011, pp 143-164.
[14] I. Goiri, W. Katsak, K. Le, T. Nguyen, and R. Bianchini. Parasol and GreenSwitch: Managing data centers powered by renewable energy, in Proc. Int. Conf. on Architectural Support for Programming Languages and Operating Systems (ASPLOS’13), 2013, pp. 51-64.
[15] S. Govindan, A. Sivasubramaniam, and B. Urgaonkar. Benefits and Limitations of Tapping into Stored Energy for Data centers, in Proc. 38th Int. Symp. on Computer Architecture (ISCA’11), 2011, pp 341-352.
[16] S. Govindan, D. Wang, A. Sivasubramaniam, and B. Urgaonkar, “Leveraging stored energy for handling power emergencies in aggressively provisioned data centers, in Proc. Int. Conf. on Architectural Support for Programming Languages and Operating Systems (ASPLOS’12), 2012 , pp. 75-86.
[17] V. Kontorinis, L. Zhang, B. Aksanli, J. Sampson, H. Homayoun, E. Pettis, T. Rosing and D. Tullsen, “Managing distributed UPS energy for effective power capping in data centers,” in Proc. 39th Int. Symp. on Computer Architecture (ISCA’12), 2012, pp. 499-499.
[18] Z. Liu, M. Lin, A. Wierman, S. Low, and L. Andrew. “Greening geographical load balancing,” in Proc. Int. Conf. on Modeling and Measurement of Computer Systems (SIGMETRICS’11), 2011, pp. 233-244.
[19] N. Deng, C. Stewart, D. Gmach, M. Arlitt, and J. Kelley, “Adaptive green hosting,” in Proc. Int. Conf. on Autonomic Computing (ICAC’12), 2012, pp. 135-144.
[20] C. Ren, D. Wang, B. Urgaonkar, and A. Sivasubramaniam, “Carbon-aware energy capacity planning for data centers,” in Proc. Int. Symp. on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS’12), 2012, pp. 391-400.
[21] P. Banerjee, C. Patel, C. Bash, A. Shah, and M. Arlitt, “Towards a net-zero data center,” ACM J. on Emerging Technologies in Computing Systems, vol. 8, issue 4, October 2012, Article No.8
[22] C. Li, W. Zhang, C. Cho, and T. Li. “SolarCore: Solar energy driven multi-core architecture power management,” in Proc. 17th Int. Symp. on High-Performance Computer Architecture (HPCA’11), 2011, pp. 205-216.
117
[23] C. Li, A. Qouneh, and T. Li, “iSwitch: Coordinating and optimizing renewable energy powered server clusters,” in Proc. 39th Int. Symp. on Computer Architecture (ISCA’12), 2012, pp. 512-523.
[24] C. Li, R. Zhou, and T. Li, “Enabling distributed generation powered sustainable high-performance data center,” in Proc. 19th Int. Symp. on High-Performance Computer Architecture (HPCA’13), 2013.
[25] C. Li, Y. Hu, R. Zhou, M. Liu, L. Liu, J. Yuan, and T. Li, “Enabling datacenter servers to scale out economically and sustainably,” in Proc. 46th Int. Symp. on Microarchitecture (MICRO’13), 2013.
[26] A. Luque and S. Hegedus. Handbook of Photovoltaic Science and Engineering, John Wiley & Sons, 2003.
[27] D. Linden and T. Reddy. Handbook of Batteries, 3rd Edition, McGraw Hill, 2002
[28] A. Jossen, J. Garche and D.U. Sauer, “Operation conditions of batteries in PV applications,” Solar Energy, Vol. 76, Issue 6, 2004, pp. 759-769.
[29] W. Kim, M.S. Gupta, G.Y. Wei and D. Brooks, “System level analysis of fast, per-core DVFS using on-chip switching regulators,” in Proc. 14th Int. Symp. on High-Performance Computer Architecture (HPCA’08), 2008.
[30] VRM Design Guideline, (April 2002), accessed on May 2014. White paper, Intel Corporation [Online]. Available: http://www.intel.com/design/pentium4/guides/249205.htm
[31] Measurement and Instrumentation Data Center, accessed on May 2014. [Online]. Available: http://www.nrel.gov/midc/
[32] D. Brooks, V. Tiwari and M. Martonosi, “Wattch: A framework for architectural-level power analysis and optimizations,” in Proc. 27th Int. Symp. on Computer Architecture (ISCA’00), 2000, pp. 83-83.
[33] N. Muralimanohar, R. Balasubramonian and N.P. Jouppi, (April 2009). CACTI 6.0: A tool to model large caches, accessed on May 2014. Technical Report HPL-2009-85, HP Labs, [Online]. Available: http://www.hpl.hp.com/techreports/2009/HPL-2009-85.html
[34] Enhanced Intel SpeedStep technology for the Intel Pentium M processor (March 2004), accessed on May 2014. White Paper, Intel Corporation, [Online] Available: ftp://download.intel.com/design/network/papers/30117401.pdf
[35] M. Patel. Wind and Solar Power Systems, CRC Press, 1999.
118
[36] D. Meisner, B. Gold and T. Wenisch, “PowerNap: Eliminating server idle power,” in Proc. Int. Conf. on Architectural Support for Programming Languages and Operating Systems (ASPLOS’09), 2009, pp. 205-216.
[37] S. Pelley, D. Meisner, P. Zandevakili, T. Wenisch and J. Underwood, “Power Routing: Dynamic power provisioning in the data center,” in Proc. Int. Conf. on Architectural Support for Programming Languages and Operating Systems (ASPLOS’10), 2010, pp. 231-242.
[38] F. Ahmad and T.N. Vijaykumar, “Joint optimization of idle and cooling power in data centers while maintaining response time,” in Proc. Int. Conf. on Architectural Support for Programming Languages and Operating Systems (ASPLOS’10), 2010, pp. 243-256.
[39] E. Pakbaznia and M. Pedram. “Minimizing data center cooling and server power costs,” in Proc. Int. Symp. on Low Power Electronics and Design (ISLPED’09), 2009, pp. 145-150.
[40] A. Gandhi, M. Harchol-Balter, R. Das and C. Lefurgy. “Optimal power allocation in server farms,” in Proc. Int. Conf. on Modeling and Measurement of Computer Systems (SIGMETRICS’09), 2009, pp. 157-168.
[41] J. Peças Lopesa,N. Hatziargyrioub, J. Mutalec, P. Djapicc, and N. Jenkinsc. “Integrating distributed generation into electric power systems: a review of drivers, challenges and opportunities,” Electric Power Systems Research, Vol. 77, Issue 9, 2007, pp. 1189-1203.
[42] M. Ton and B. Fortenbery, (March 2008). DC power for improved data center efficiency, accessed on May 2014. Lawrence Berkeley National Laboratory, Technical Report, [Online]. Available: http://hightech.lbl.gov/documents/data_centers/dcdemofinalreport.pdf
[43] X. Fan, W. Weber and L. Barroso, “Power provisioning for a warehouse-sized computer,” in Proc. 34th Int. Symp. on Computer Architecture (ISCA’07), 2007, pp. 13-23.
[44] X. Wang, M. Chen, C. Lefurgy, and T. Keller. “SHIP: Scalable hierarchical power control for large-scale data centers, in Proc. 18th Int. Conf. on Parallel Architecture and Compilation Techniques (PACT’09), 2009, pp. 91-100.
[45] SPECPower Results, accessed on May 2014. [Online]. Available: http://www.spec.org/power_ssj2008
[46] Logs of Real Parallel Workloads from Production Systems, accessed on May 2014. [Online]. Available: http://www.cs.huji.ac.il/labs/parallel/workload/logs.html
119
[47] Western Wind and Solar Integration Study, accessed on May 2014. [Online]. Available: http://www.nrel.gov/wind/systemsintegration/wwsis.html
[48] V. Soundararajan and J. Anderson, “The impact of management operations on the virtualized data center,” in Proc. 37th Int. Symp. on Computer Architecture (ISCA’10), 2010, pp. 326-337.
[49] M. Zhao and R. Figueiredo, “Experimental study of virtual machine migration in support of reservation of cluster resources,” in Proc. Int. Workshop on Virtualization Technology in Distributed Computing, 2007, Article No.5
[50] R. Lasseter and P. Piagi. “Microgrid: A conceptual Solution,” in Proc. IEEE Annual Power Electronics Specialists Conference, 2004, pp. 4285-4290.
[51] The role of distributed generation and combined heat and power (CHP) systems in data centers, (August 2007), accessed on May 2014. Technical Report, EPA, [Online]. Available: http://www.epa.gov/chp/documents/datactr_whitepaper.pdf
[52] B. Kirby and E. Hirst, (January 2000). Customer-specific metrics for the regulation and load-following ancillary services, accessed on May 2014. Technical Report, Oak Ridge National Lab, [Online]. Available: http://energy.lbl.gov/ea/certs/pdf/ornl-con-474.pdf
[53] S. Chowdhury and P. Crossley. Microgrid and active distribution networks, The Institute of Engineering and Technology, 2009
[54] P. Piagi and R. Lasseter. “Autonomous control of microgrids,” in Proc. IEEE Power Engineering Society General Meeting, 2006, pp. 8-8
[55] HOMER Energy Modeling Software, http://homerenergy.com/
[56] M. Stansberry and J. Kudritzki, (April 2012). The 2012 Uptime Institute data center industry survey, accessed on May 2014. Write Paper, The Uptime Institute, [Online]. Available: http://uptimeinstitute.com/images/stories/Uptime_Institute_2012_Data_Industry_Survey.pdf
[57] D. Bouley, (2011). Estimating a data center’s electrical carbon footprint, accessed on May 2014. White Paper #66, Schneider Electric, [Online]. Available: http://www.apcmedia.com/salestools/DBOY-7EVHLH/DBOY-7EVHLH_R0_EN.pdf
[58] P. Sarti, (April, 2011). Battery cabinet hardware v1.0, accessed on May 2014. Write Paper, Facebook Open Compute Project, [Online], Available: http://www.opencompute.org/wp/wp-content/uploads/2011/07/DataCenter-Battery-Cabinet-Specifications.pdf
120
[59] H. Sher, and K. Addoweesh, “Micro-inverters - promising solutions in solar photovoltaics,” Energy for Sustainable Development, Elsevier, 2012, pp. 389-400
[60] National survey on data center outages, (September 2010), accessed on May 2014. White Paper, Ponemon Institute, [Online]. Available: http://www.emersonnetworkpower.com/documentation/en-us/brands/liebert/documents/white%20papers/data-center-optimization_24655-r09-10.pdf
[61] Modbus Protocol, http://www.modbus.org/
[62] H. Bindner, T. Cronin, P. Lundsager, J. Manwell, U. Abdulwahid, I. Gould, (April 2005). Lifetime modelling of lead acid batteries, accessed on May 2014. Risø National Laboratory, Technical Report, [Online]. Available: http://orbit.dtu.dk/fedora/objects/orbit:88309/datastreams/file_7710966/content
[63] D. Doerffel, and S. Sharkh. “A critical review of using the Peukert equation for determining the remaining capacity of lead-acid and Lithium-ion batteries.” Journal of Power Sources, vol. 155, issue 2, 2006, pp. 395-400.
[64] S. Huang, J. Huang, J. Dai, T. Xie, and B. Huang, “The HiBench benchmark suite: Characterization of the MapReduce-based data analysis. Data Engineering Workshops, IEEE Int. Conf. on Data Engineering, 2010, pp. 41-51.
[65] CloudSuite 2.0, accessed on May 2014. [Online]. Available: http://parsa.epfl.ch/cloudsuite
[66] Watts Up? Meters, https://www.wattsupmeters.com
[67] D. Feldman, G. Barbose, R. Margolis, R. Wiser, N. Darghouth, and A Goodrich, (November 2012). Photovoltaic (PV) pricing trends: historical, recent, and near-term projections, accessed on May 2014. Technical Report, National Renewable Energy Laboratory and Lawrence Berkeley National Laboratory, [Online]. Available: http://www.nrel.gov/docs/fy13osti/56776.pdf
[68] Navigant Consulting, (January 2006). A review of PV inverter technology cost and performance projections, accessed on May 2014. National Renewable Energy Lab, Technical Report, [Online]. Available: http://www.nrel.gov/docs/fy06osti/38771.pdf
[69] R. Simpson, (February 2012). Levelized cost of energy from residential to large scale PV, accessed on May 2014. in Proc. the Applied Power Electronics Conference and Exposition, [Online]. Available: http://www.psma.com/sites/default/files/uploads/tech-forums-alternative-energy/presentations/2012-apec-256-evolution-photovoltaic-inverter.pdf
121
[70] SRRL Baseline Measurement System, accessed on May 2014. [Online]. http://www.nrel.gov/midc/srrl_bms/
[71] G. Burch, (August 2001). Hybrid renewable energy systems, accessed on May 2014. in Proc. The Natural Gas / Renewable Energy Workshops, U.S. Department of Energy, [Online]. Available: http://www.netl.doe.gov/publications/proceedings/01/hybrids/Gary%20Burch%208.21.01.pdf
[72] Information optimization: Harness the power of big data, (December 2012), accessed on May 2014. White Paper, Hewlett-Packard, [Online]. Available: http://www.hp.com/hpinfo/newsroom/press_kits/2012/HPDiscoverFrankfurt2012/IO_Whitepaper_Harness_the_Power_of_Big_Data.pdf
[73] W. Johnston, E. Dart, M. Ernst, and B. Tierney, (2003). Enabling high throughput in widely distributed data management and analysis systems: lessons from the LHC, accessed on May 2014. Technical Report, Lawrence Berkeley National Laboratory, [Online]. Available: http://www.es.net/assets/pubs_presos/High-throughput-lessons-from-the-LHC-experience.Johnston.TNC2013.pdf
[74] C. Li, R. Wang, T. Li, J. Yuan, and D. Qian, “Managing green datacenters powered by hybrid renewable energy systems,” in Proc. Int. Conf. on Autonomic Computing, 2014.
[75] S. McCluer, (2011). Battery technology for data centers and network rooms: lead-acid battery options, accessed on May 2014. White Paper #30, APC, [Online]. Available: http://www.apcmedia.com/salestools/SADE-5TNQZD/SADE-5TNQZD_R12_EN.pdf
[76] F. Zhang, S. Lasluisa, T. Jin, I. Rodero, H. Bui, and M. Parashar, “In-situ feature-based objects tracking for large-scale scientific simulations,” SC Companion: High Performance Computing, Networking Storage and Analysis (SCC), 2012, pp. 736-740.
[77] S. Lakshminarasimhan, D. Boyuka, S. Pendse, X. Zou, J. Jenkins, V. Vishwanath, M. Papka, and N. Samatova, “Scalable in-situ scientific data encoding for analytical query processing,” in Proc. int. Symp. on High-Performance Parallel and Distributed Computing (HPDC’13), 2013
[78] B. Aksanli, P. Eddie, and R. Tajana, “Architecting efficient peak power shaving using batteries in data centers,” in Proc. IEEE Int. Symp. on Modelling, Analysis & Simulation of Computer and Telecommunication Systems (MASCOTS’13), 2013, pp. 242-253.
122
[79] M. Ebbers, A. Galea, M. Schaefer, M. Khiem, (August 2008). The green data center: Steps for the journey, accessed on May 2014. White Paper, IBM, [Online]. Available: ftp://ftp.boulder.ibm.com/software/uk/itsolutions/optimiseit/energy-efficiency-solutions/2-new-enterprise-data-centre-the-green-new-enterprise-data-centre-red-books.pdf
[80] Z. Kishimoto, (August, 2013). IT controls power at data centers, accessed on May 2014. Technical Blog, AltaTerra Research Network, [Online]. Available: http://www.altaterra.net/blogpost/587092/168544/IT-Controls-Power-at-Data-Centers--part-1
[81] Environmental impacts of transmission lines, accessed on May 2014. White Paper, Public Service Commission of Wisconsin, [Online]. Available: http://psc.wi.gov/thelibrary/publications/electric/electric10.pdf
[82] J. Glanz, (September 2012). Power, pollution and the internet, accessed on May 2014. [Online]. Available: http://www.nytimes.com/2012/09/23/technology/data-centers-waste-vast-amounts-of-energy-belying-industry-image.html?pagewanted=all
[83] C. Bange, T. Grosser, and N. Janoschek, (February 2013). Big data survey Europe: Usage, technology and budgets in European best-practice companies, accessed on May 2014. White Paper, BARC, [Online]. Available: http://www.pmone.com/fileadmin/user_upload/doc/study/BARC_BIG_DATA_SURVEY_EN_final.pdf
[84] N. Femia, G. Petrone, G. Spagnuolo, and M. Vitelli. “Optimization of perturb and observe maximum power point tracking method”, IEEE Transactions on Power Electronics, Vol. 20, No. 4, July 2005, pp. 963-973.
123
BIOGRAPHICAL SKETCH
Chao Li was born and raised in Taiyuan, Shanxi, China. He earned his B.S. (with
Chu Kochen Certificate of Honors) in electronic and information engineering from
Zhejiang University, China. He was a visiting student at UCLA in August 2008.
Chao joined the Intelligent Design of Efficient Architectures Laboratory (IDEAL)
at the University of Florida in August 2009. His research was generously supported by
several prestigious awards, including a four-year UF Alumni Graduate Fellowship Award
from 2009 to 2013, a Yahoo! Key Scientific Challenges Program Award in Green
Computing in 2012, and a Facebook Graduate Fellowship in Computer Architecture in
2013. He was a winner of the IEEE HPCA 2011 Best Paper Award.
Chao was a Ph.D. recipient of the 2010 University of Florida Gator Engineer
Recognition Award for proposing sustainable solutions to the computer energy
challenge and global climate change. He was named an Outstanding International
Student by the UF International Center in 2010 and 2013. He received the ACM A.M.
Turing Centenary Celebration Student Scholarship in 2012 and attended the once-in-a-
lifetime ceremony with 34 Turing Award winners participating in the event. He was also
awarded the prestigious 2013 Chinese Government Award for Outstanding Self-
Financed Students Studying Abroad from the China Scholarship Council (CSC).
During his graduate studies, Chao served as the organizing committee member
for the 2013 Annual IEEE International Conference on Networking, Architecture, and
Storage (NAS) and the 2014 Annual IEEE International Symposium for High-
Performance Computer Architecture (HPCA). He was a student member of IEEE, IEEE
Computer Society, ACM, and ACM SIGARCH.