a cyber physical approach to a combined hardware-software
DESCRIPTION
Presentation by Josué Pagán at DCIS 2013 conference, organized by CEIT (Nov 27th, 2013)TRANSCRIPT
A Cyber Physical Approach to
Combined HW-SW Monitoring for
Improving Energy Efficiency in
Data Centers
Josué Pagán, Marina Zapater, Oscar Cubo, Patricia Arroba, Vicente Martín and José M. Moya
Universidad Politécnica de Madrid1 / 20
Contents
1. Power consumption problem in Data Centers
I. Introduction
II. Related work
2. Optimization Framework & Data analysis
I. A Cyber-physical system
II. Data analysis and sensor configuration
3. Results
4. Conclusions
2
1. POWER CONSUMPTION
PROBLEMS IN DATA CENTERS
Data Center. Consumption
3
1. Power consumption
problems in Data Centers
• The numbers of the energy problem:– DC world power consumption >1.3%
– In urban areas >50% of DC exceeds power grid capacity
– USA: 80 TWh/year in 2011 = 1.5 x NY
•
4Projection of total electricity use by datacenters in the US and the world based on Koomey’s and EPA’s data
Power >600 TWhr expected in 2015 in the global footprint
Data Centers’ power consumption is unsustainable
5
Cooling
• Allow the room temperature to increase
• Longer task → cooler server
Computation
• Balancing workload between servers
• Reducing voltage/ frequency (DVFS)
Holistic(IT+cooling)
• Room environment affects (environmental monitoring)
• Measuring server, workload and environmental variables to improve energy efficiency → usage of a CPS
1. Power consumption
problems in Data Centers
• Related work (approaches)
– These two approaches are not enough individually
1. Consumption problems in
Data Centers
6
Requirements Our contributions
Energy optimization
Make a holistic optimization framework including environmental,
server and workload information
Dynamically adapt on runtime to workload and environment
Gather, monitor and analyze in real time
Gather useful data at the appropriate rate
In a non-intrusive way, reducing the data collected with an adaptable
sampling rate
2. OPTIMIZATION FRAMEWORK
& DATA ANALYSIS
Cyber-Physical System. Data acquisition
7
• One step ahead. Optimization– 80% Wpeak – 30% of workload (↓η)
– An energy model supposes apply optimizations over the Data Center
2. Optimization Framework &
Data analysis
8
GATHERDATA
GENERATEKNOWLEDGE
PROPOSEOPTIMIZATIONS
• Monitoring– How a Data Center works?
– 30-50% cooling→ energy optimization 9
2. Optimization Framework &
Data analysis
• What measure and why– Environmental monitoring
Inlet and outlet temperature
Differential pressure
10
2. Optimization Framework &
Data analysis
– Server monitoringServer consumption, CPU temperature, fan speed
• …to predict
• How…
• exploring sampling intervals
2. Optimization Framework &
Data analysis
11
– Different sampling rates for different parameters
– Temperature and power values
for AMD server under the
benchmark SPEC CPU 2006
• Using… Multilevel star topology architecture
12
Air conditioning- Exhaust
temperature, RH% and airflow
WSN- Reconfigurable low -power:
only useful data without information loss
- Adapt to changes in the environment
RM- Spatio-temporal allocation
- Possibly to change decisions if needed
Server Sensors- Internal sensors
- Polled via SW
Gateway-Fan-less, managed with a light OS
-Receive, store, analyze and convert data. Establishes a timestamp.
-Sends data to the opt. platform
2. Optimization Framework &
Data analysis
3. RESULTS
13
3. Results
14
WSN deployment• Applied over Magerit Supercomputer in CeSViMa Supercomputing and Visualization Center
of Madrid• Cluster 9 racks 260 servers
3. Results
15
– The goal: develop techniques to allow energy optimization in real environments
– With reconfigurable sampling rate:
– we achieve up to 68 % of reduction in gathered data
– Increase the WSN’ s life time depending on the occupancy
4. CONCLUSIONS
16
4. Conclusions
17
Energy efficiency has to be faced in a holistic way
We propose an optimization framework monitoring environmental, server and workload parameters
After a first monitoring study: a WSN has been deployed to gather environmental data
Up to a 68% of reduction in the amount of gathered data
Maximizing the life time of WSN nodes
Solution applied in a real case study
FIN
This project has been funded in part by the INNPACTO \ LPCLOUD: "Optimal Management Of low-power modes in cloud computing" IPT-2012-1041-430000, developed in collaboration with Elite Ermestel and Converging Technologies and the CDTI project \ CALEO:
Distribution of operational thermal and optimization of energy consumption in data centers, "developed in collaboration with INCOTEC. The author gratefully acknowledges the computer resources, technical expertise and assistance provided by the Supercomputing and
Visualization Center of Madrid (CeSViMa). 18
Thank you for your attention
4. Results and Conclusions
19
• Results: gathering data• Inlet and outlet temperature
Magerit Supercomputer
20
• Cluster 9 racks 260 servers 245 are IBM PS702 2S
o 16 Power7 processors @ 3.3 GHz
o 32 GB of RAM
15 are IBM HS22
o 8 Intel Xeon processors @ 2.5 GHz
o 96 GB of RAM
200 TB of storage
21
Industry
Software
22
• Pasarela
Cyber-Physical System
23
Psychrometric chart
24
Differential pressure and
airflow
25
Installing nodes [6T+1H]
26