planetary scale computing towards planetary scale...

June 2003- 1

Planetary Scale

Computing

© 2001-2003 Hewlett Packard Company

Towards Planetary Scale Computing

next generationinternet computing

Rich Friedrichdirector

internet systems and storage labhp laboratories

June 2003http://www.hpl.hp.com/research/internet/

June 2003- 2

Planetary Scale

Computing


Outline

● The rise of the Internet Data Center● Why scale to the planet?● Planetary Scale Computing at HP Labs● From research to reality● A research platform

June 2003- 3

Planetary Scale

Computing


The rise of the internet data center

● Confluence of web based distributed applications, Linux based rack servers, and internet data centers has enabled applications unimaginable a decade ago.

● However, these new environments are problematic: ■ dedicating hardware to specific applications limits flexibility,■ varying application demands result in poor server utilization, ■ rising complexity escalates operational costs, and■ emerging applications will consume 10-100X more resources■ IPv6 world with 2128 nodes ranging from nanobots to 128 way

servers ■ increasing server density generates energy and cooling issues

! Consequently, a new conceptual model for large-scale computing is required that addresses flexibility, utilization and cost while providing performance, security and fault isolation.

June 2003- 4

Planetary Scale

Computing


Why IT services on a planetary scale?

● Centralization is the bane of tomorrow■ Simple scaling arguments dictate that “machine rooms” will

be dispersed■ Ownership of academic, business and consumer content will

be dispersed, but not necessarily public■ A sea of interconnected resources now exists■ Rich media from the masses

● Geographic dispersion of virtual teams■ Few people only interact with colleagues in the next office■ Workload demand follows the sun■ Supply chains■ 50% of professionals in 2006 may telecommute■ Online entertainment

! Consequently, new conceptual model must take into account large scale distributed services

June 2003- 5

Planetary Scale

Computing


planetary scale computing

vision

● We envision a world where distributed services execute on a utility that dynamically and securely allocates globally connected resources on demand

● …a global commercial GRID

● …will do for resources what the Web did for documents

■ provide uniform, ubiquitous access to globally connected server and storage resources

■ provide infrastructure on demand

■ eliminate resource shortages

June 2003- 6

Planetary Scale

Computing


planetary scale computing

requirements

● Large-scale and federation● Self-adapting for varying workload demands and resource capacity

● Resilient in the face of failure or attack

● Trust and privacy● Dynamic and evolutionary events

● Self-describing, verifiable policy model

● Mobile clients and storage● O(105) elements per data center

● Economical to deploy and operate

● Supporting emerging 21st

century applications such as rich media, bioinformatics, massive personalization, sensor networks

June 2003- 7

Planetary Scale

Computing


“Programmable Data Center”

Distributed, Shared Resources:

SecureFlexibleDynamicTraded

thesis: next generation of computing: the data center is the computer

“Internet Data Center”

Dedicated Resources:SecureInflexibleStatic

2001

Supplychain

SalesHR

R&D Mfg

Supplychain

SalesHRR&DMfg

2007

June 2003- 8

Planetary Scale

Computing


internet systems

architecture research

the data center is the computer

● What high density, low power, high performance computing architectures most economically support pervasive computing?

● What are the simple building blocks of processing, communications, storage and power that support dynamic allocation of virtualized resources?

● Where should processing occur? Where should data reside?

● Can ethernet provide a single server/storage fabric?

● What is the science of large scale dynamical systems that estimates probabilistic behavior based on small scale experiments?

June 2003- 9

Planetary Scale

Computing


“Programmable Data Center”

Distributed, Shared Resources:

SecureFlexibleDynamicTraded

thesis: programmable data center requires a program: the data center OS

Ford HPP&GYahooebay

2007“Data Center OS”

Automatically allocate distributed, shared resources:

SecureFlexibleDynamicFederationSelf-adapting

“Data Center OS”

Automatically allocate distributed, shared resources:

SecureFlexibleDynamicFederationSelf-adapting

June 2003- 10

Planetary Scale

Computing


automated resource control

research

the data center OS

! All physical resources controllable; i.e. servers, routers, storage and energy.

● What decentralized mechanisms support distributed resource allocation?

● What automated reasoning systems can eliminate the complexity of controlling large scale systems?

● What sensors and actuators are necessary?

● What control techniques are applicable to reactive and predictive events?

● What time scales are appropriate for control?

● How is security enforced?● How are control measures

and decisions coordinated across federated systems?

June 2003- 11

Planetary Scale

Computing


basic idea: consumer/supplier

applications

virtualized resource pool

Demandintelligent brokering:matchservice demandwithresource capacity

Supply

intelligent provisioning:effective use of physical resources

June 2003- 12

Planetary Scale

Computing


Early insights: Internet workload characterization and control

Applications

virtualized resource pool

Supply Demand

User Demand for a Fortune 100 WWW Server

120000

160000

200000

240000

280000

Req

uest

s pe

r Day

Web Request Size Distribution

0

20000

40000

60000

80000

100000

120000

2 4 8 16 32 64 128

256

512

1024

2048

4096

8192

1638

4

3276

8

6553

6

1310

72

2621

44

5242

88

1048

576

2097

152

4194

304

8388

608

1677

7216

Request Size (bytes)

Freq

uenc

y

Fortune 100 daily demand web pages ..

… file size evolution …

OL

TP

E-C

omm

Web

97

Web

2001

New media types:

voice, audio, video

Growth 10-15% per month3x in 1 year10x in 2 years

characterization

M. Arlitt, D. Krishnamurthy, J. Rolia, "Characterizing the Scalability of a Large Web-based Shopping System," ACM Transactions on Internet Technology, June 2001.

Server QoSadaptable control

● Offered Load = 300% (at t=13)● Desired utilization = 85%

Nina Bhatti and Rich Friedrich, "Web Server Support for Tiered Services" IEEE Network, 13(5):64-71, September 1999.

June 2003- 13

Planetary Scale

Computing


• For the six data centers we looked at:• More than 50% of servers have a utilization <= 10%; 85% are <= 25% • All the data centers have servers that are very busy• The identity of the servers that are busy varies with time

Data center server utilization

June 2003- 14

Planetary Scale

Computing


A reference architecture frameworkfor a data center OS

Non-stoptrusted utility

Real time data msmt & mining

IntelligentResourcecontrol

Programmable resource utility

Resource management system

Service Spec &deployment system

GRID Services(OGSI)

UDC UDC-lite horizontal-scale cluster SMP blades energy cooling trust privacy

Business processes

BAM Business applications& web services (SLAs)

Non-stoptrustedservices

automating operational processes based on business priorities

June 2003- 15

Planetary Scale

Computing


Adaptive internet data center circa 2000

accesstier

webtier

applicationtier

databasetier

edge routers

routingswitchesauthentication, DNS,

intrusion detect, VPNweb cache

1st level firewall

2nd level firewall

load balancingswitches

web serversweb page storage

(NAS)

databaseSQL servers

storage areanetwork(SAN)

applicationserversfiles

(NAS)

switches

switches

switchedfabric

processingelements

storageelements

infrastructure on demand

internet

intranet

wire infrastructure once...

rewire programmatically,

dynamically re-provision resources

J. Rolia, S. Singhal, R. Friedrich, “Adaptive Internet Data Centers,” SSGRR 2000, European Computer and

eBusiness Conference, L’Aquila, Italy, July 2000

An architecture for tomorrow

Programmable resource utilityTrusted

resources



UDC UDC-lite horizontal-scale cluster SMP blades energy cooling

Real timedata mining (ZLE)

Non-stop

trusted utility

IntelligentResource

control

GRIDServices(OGSA)



Business processes

June 2003- 16

Planetary Scale

Computing


From research to reality

● HP announced the Utility Data Center (UDC) Nov 2001

■ incorporates key HP Labs concepts and technologies

● Based on HP Labs research on adaptive internet data center

■ ability to direct resources to any application dynamically

■ self healing, policy driven.■ Heterogeneous environments:

Windows, Linux, HP-UX, Sun Solaris, Cisco, ProCurve, EMC

switchedfabric

processingelements

storageelements

infrastructure on demand

internet

intranet

… to create a dynamically configurable utility fabric that

can be programmed per service or customer, based on

SLAs and demand…


resources





Non-stop

trusted utility

IntelligentResource

control

GRIDServices(OGSA)



Business processes

June 2003- 17

Planetary Scale

Computing


creating a service with the UDC

1. Architect new service:

LB

WEB WEB WEB

APP

FWSvc “A”

Svc “A”

2. Build a service template:

APP

FW 1U Linux

2U NT

HP-UX

applianceLB

WEB WEB WEB

• Install apps

3. Ignite the service

Free

Discover and apply free resources

• Specify connectivity• Auto-configure

network and storage• Auto-load OSes

Svc “A”

June 2003- 18

Planetary Scale

Computing


IT Service ConsumerBusiness Impact Analysis2

2Business Process Cockpit; M. Sayal, F.Casati, U.Dayal, M.C.Shan, VLDB 2002

IT Service ProviderResources

Linking service management and business processes

ServiceSLO/SLA

ServiceSLO/SLA

ServiceSLO/SLA

ServiceSLO/SLA

IT manager

1Towards regulating electronic communities with contract; M.Morciniec, M.Salle, B.Monahan, ICAIL 2001

CIO

Management by Contract1

LOB manager

June 2003- 19

Planetary Scale

Computing


SmartFrog: service description and deployment

• Flexible configuration description language

• precise, desired configuration of applications composed of sets of components running across a distributed system

• declarative, prototype-based• predicates, constraints

• Service deployment architecture for massive systems

• realize application description• monitor and manage the resulting

applications through their lifecycles• No single point of control

applicationapplicationapplicationapplicationdescriptiondescriptiondescriptiondescription

SmartFrogSmartFrogSmartFrogSmartFrog distributed distributed distributed distributed deployment systemdeployment systemdeployment systemdeployment system

SmartFrogSmartFrogSmartFrogSmartFrog notationnotationnotationnotation

realizes running, realizes running, realizes running, realizes running, distributed distributed distributed distributed

applicationsapplicationsapplicationsapplications

managed, managed, managed, managed, monitored through monitored through monitored through monitored through

lifecyclelifecyclelifecyclelifecycle

• which application which application which application which application components?components?components?components?

• running where?running where?running where?running where?• how is each component how is each component how is each component how is each component

configured?configured?configured?configured?• how are the component how are the component how are the component how are the component

lifecycles sequenced?lifecycles sequenced?lifecycles sequenced?lifecycles sequenced?• how are components how are components how are components how are components

related?related?related?related?

Patrick Goldsack, "SmartFrog: A framework for configuration", from the Workshop on Large-Scale System Configuration, Edinburgh, November 2001. (Online proceedings available at www.dcs.ed.ac.uk/home/paul/wshop)

Managing the sea of software versions and dependencies


resources





Non-stop

trusted utility

IntelligentResource

control

GRIDServices(OGSA)



Business processes

June 2003- 20

Planetary Scale

Computing


CPU allocations

Case Study- information on cpuutilization of 48 servers in a data center- simulated traces of cpuusage based on profiles from raw data

4 8 12 16 20 240

100

200

300

400 Static (util=0.5)

Hour of Day 4 8 12 16 20 24

0

100

200

300

400 Guaranteed (util=0.56)

Hour of Day

4 8 12 16 20 240

100

200

300

400Prob Best Effort (util=0.87)

Hour of Day

p=0.999

4 8 12 16 20 240

100

200

300

400Prob Best Effort (util=0.95)

Hour of Day

p=0.99

Resource Access Management for Computing Utilities

• Enterprise applications: continuously available, demands vary with time and user request loads, high peak-to-mean ratio

• Adaptive infrastructure: programmable data centers that offer shared resources

• Our focus: increase asset utilization and enable QoS for resource access

• Capacity planning/admission control/resource allocation --- use demand profiles, exploit time based allocation and statistical multiplexing

• Class of Service (CoS) --- static, guaranteed interval based access, best effort, probabilistic best effort: offers resources on demand with a specific probability p

Publications- J. Rolia, X. Zhu, M. Arlitt and A. Andrzejak, “Statistical service assurance for applications in utility Grid environments,” MASCOTS 2002. - J. Rolia, X. Zhu and M. Arlitt, “Resource access management for a utility hosting enterprise applications,” to appear at IM 2003.


resources





Non-stop

trusted utility

IntelligentResource

control

GRIDServices(OGSA)



Business processes

June 2003- 21

Planetary Scale

Computing


0

2000

4000

6000

8000

10000

12000

14000

0 100 200 300 400 500 600 700 800 900

Nu

mb

er

of

Se

ssio

ns/D

ay

Day Number

HPC

Sessions

dynamics of media sites

how do we deal with bursty loads?

1 1.2 1.4 1.6 1.8 2Risk

0.1

0.12

0.14

0.16

0.18

0.2

nr

ut

eR

Computation as economicsUsing dynamic portfolio approaches

how do we manage risk and return?

market components: computational analogs:

resources hardware and software

agents programs’ choices

preferences computational needs

Bernardo Huberman: journal of econ. dynamics and control, 22, 1169 (1998)

June 2003- 22

Planetary Scale

Computing


Turning the UDC into a “power station” for the commercial GRID

● Use an offline tool to design application topologies

● Use the Globus toolkit to■ submit resource

requests to a UDC■ create a “farm” (a multi-

tier topology) request■ return the access

information for the farmto the user

● UDC provides secure, dynamically allocatableresources for the GRID

Sven Graupner, Jim Pruyne, Sharad Singhal, Making the Utility Data Center a Power Station on the Commercial Grid, GlobusWorld 2003.

June 2003- 23

Planetary Scale

Computing


Scalable commodity open source platform

http://www.gelato.org


resources





Non-stop

trusted utility

IntelligentResource

control

GRIDServices(OGSA)



Business processes

• Automated application optimization• Secure kernel platform• stack unwind and perfmon

June 2003- 24

Planetary Scale

Computing


High Performance Computing

Dual Processor PEAK Bytes/FLOPSmall is BAD

-

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1H03 2H03 1H04 2H04 1H05 2H05

IA-32IA-64

Godiva

June 2003- 25

Planetary Scale

Computing


Appia: automated SAN fabric designThe problem: Given flow requirements between hosts and storage devices, find a minimum cost reliable network to support all flow requirements simultaneously.

hosts

storage devices

flow requirements

(MB/s)

links& ports

fabric nodes (hubs & switches)

& a path for each flow

Ward, J., O'Sullivan, M., Shahoumian, T., Wilkes, J. Appia: automatic storage area network fabric design. File and Storage Technologies (FAST) Conference, 2002.

The solution: a software tool that produces provably correct, reliable and cost-effective designs in minutes.

3 days$4m

10 mins$1.4m


resources





Non-stop

trusted utility

IntelligentResource

control

GRIDServices(OGSA)



Business processes

June 2003- 26

Planetary Scale

Computing


Federated Array of Bricks (FAB):building an affordable and scalable disk array

● Goals■ drop in replacement for

disk array■ “infinite” scale■ incrementally upgradeable■ self-managing: no people

● Benefits■ smooth, incremental

capacity+performance scaling

■ enterprise-class reliability, functionality, management

■ low entry price■ best-in class absolute

price

Clients (typically servers, blades, etc)

storage bricks

iSCSI, FC, SCSI, SAS, …(read and write)(direct connect or SAN)

FAB protocols

Back End Network – Ethernet+RDMA, … (our choice)

Research: quorum-basedreplication scheme, dynamicload balancing and online

reconfiguration

Ref: “FAB: enterprise storage systems on a shoestring”, HotOS 2003

June 2003- 27

Planetary Scale

Computing


Hippodrome: automatic storage management

Hippodrome: running circles around storage administration Eric Anderson, Michael Hobbs, Kimberly Keeton, Susan Spence, Mustafa Uysal, and Alistair Veitch. Conference on File and Storage Technology (FAST'02) January 2002

AnalyzeAnalyzeAnalyzeAnalyzeworkloadworkloadworkloadworkloadAnalyzeAnalyzeAnalyzeAnalyze

workloadworkloadworkloadworkloadImplement Implement Implement Implement

designdesigndesigndesignImplement Implement Implement Implement

designdesigndesigndesign

Design newDesign newDesign newDesign newsystemsystemsystemsystem

Design newDesign newDesign newDesign newsystemsystemsystemsystem

Design system to meet workload requirements

Configure devices & migrate data

Learn workload performance characteristics

benefit: lower operational cost for storage


resources





Non-stop

trusted utility

IntelligentResource

control

GRIDServices(OGSA)



Business processes

June 2003- 28

Planetary Scale

Computing


Corporate Immune Systems

Time

ResponseIn

fect

ed M

achi

nes

Problems

Response

Prevention

Speed

Fast

Slow

ResilientInfrastructure

ResilientInfrastructure

Virus ThrottlingSlowing down the spread by throttleing new connections

Matthew M. Williamson "Throtting viruses: restricting propagation to defeat malicious mobile code", proceedings of ACSAC

conference 2002, Las Vegas, NV

June 2003- 29

Planetary Scale

Computing


• Microprocessor Power – 100 W (tenfold growth in ten years)

• Microprocessor Power Density - 200 W/cm2 (by 2003, today 60 W/cm2)

• High System Power Density – 300 W, thin 1U form factor

• High EIA Rack Power Density – 10 to 15 KW per EIA Rack foot print

• High Room Power Density - 2700 W/m2 (~300 W/ft2, today at 70 W/ft2)

Energy to Remove Heat

.100 KW 1000+ KW10 - 15 KW

500 KW1 KW0.005 KW

Heat Generated

FlowThermo-dynamics

Thermal Challenges from Chips to Data Centers

June 2003- 30

Planetary Scale

Computing


HeatExchanger

pump

• Precise spray commensurate with the heat load on the chip

Evaporative Spray CoolingConductive Interface in high power processing and communication devices (>200 W/cm2, total power of 75 W) will not work

Why? How?

Inkjet assisted spray cooling

June 2003- 31

Planetary Scale

Computing


Dynamic thermal management in large scale data centers

Patel, C.D., Sharma, R.K, Bash, C.E., Beitelmal, A, Thermal Considerations in Cooling Large Scale High Compute Density Data Centers, ITherm 2002 – 8th Intersociety Conference on Thermal and Thermomechanical Phenomena in Electronic Systems” May 2002, San Diego, California

• Power Density becoming critical -

• Microprocessor: 200 W/cm2

(by 2003, today 60 W/cm2)

• System – 300 W, thin 1U form factor 10 to 15 KW per EIA Rack foot print

• Room- 2700 W/m2 (~300 W/ft2)

• Affects reliability and cost

! Use 3D modeling to understand thermal characteristics of data centers

! Sensor networks and robotic in situmeasurements

! Exploit this for dynamic resource allocation and proper provisioning


resources





Non-stop

trusted utility

IntelligentResource

control

GRIDServices(OGSA)



Business processes

June 2003- 32

Planetary Scale

Computing


Summary

● HP Labs believes systems research needs to address

■ A new conceptual model for large scale internet computing and storage built upon commodity components and resource virtualization

■ Utility computing architectures, mechanisms and policies

■ Intelligent, decentralized resource control to simplify operation and reduce lifecycle costs

■ Pervasive, unified security model to ensure privacy and mitigate denial of service attacks

June 2003- 33

Planetary Scale

Computing


Twin UDCs in HP Labs● Built the first large Planetary

Computing infrastructure in Palo Alto (US) and Bristol (UK)

■ Learn what it takes to build a solution■ Move HPL IT services to the UDC

● The first Virtualized Data Center ■ From Server, storage, networks to energy management

● Platform for HP experimentation● Raise the level of research platform

collaboration■ HPL, Academia, Partners■ Interested?" [email protected]

planetary scale computing towards planetary scale...

Documents