workshop on green and environmental computing - helsinki

Workshop on Green and Environmental Computingp g27th – 28th October 2010

CSC – IT Center for Science Espoo, Finland

Middl h Middleware research for green Data Centers

Prof. Jordi TorresBarcelona Supercomputing Center (BSC)UPC Barcelona TechUPC – Barcelona Tech

1Jordi Torres – Barcelona Supercomputing Center (BSC) 2010 e-InfraNet Workshop on Green and Environmental Computing

Contents

Motivation for reduction of energy consumption

What can we do for DC power efficiency?

Power aware Resource Management

Research at a Glance

Conclusions


The motivations are well-known by all of youby all of you

Impact of computing on the environment:

ICT industry produces roughly 2 to 3 percent of global


ICT industry produces roughly 2 to 3 percent of global GHG emissions.

And we can’t forget …g

Opex Data Center 5 years (%)

Data courtesy of AST IT Infrastructures

Calculation for avg density 7kw/rackInfrastructure amortization counted as OpexHW and SW excluded


(www.AST-global.com)HW and SW excludedData Center Infrastructure valued as Opex

Goal: Reduce DC energy usegy

Has become a first-order objective in the design and operation of theoperation of the Data Centers

Source: EPA report to Congress 2007

How can we reduce it?


Contents





Conclusions


What can we do for DC power efficiency?efficiency?

Can we Minimize losses and thermal/coolingCan we Minimize losses and thermal/cooling overheads in a Data Center?Example: Reducing cooling costsExample: Reducing cooling costs“Best Practices to save energy at Marenostrum”

Source: “El reto operacional de dirigir el tercer supercomputador supercomputador más grande de Europa” Sergi Girona, Director Operaciones BSC-CNS, 2009.


Foto: BSC

Best practice at Marenostrum

Problem measured:Problem measured: to much under-floor pressureTest: Move some floor tiles

Benefits observed: – Improvement of AC

equipmentequipment performance

– Improved the rack-


pbottom temperature

Foto: BSC


Substituting Floor Tiles– Composite

• 20% opening

– Metallic tile• 40% opening40% opening

B fit b dBenefits observed:– Less working pressure for

the Cooling components– All bladecenters reduced by

2º C– Cold barrier that prevents


Cold barrier that preventsthe reflux of hot air

Foto: BSC


Problem measured : not all the Racks have the same temperature

Proposal:Methacrylate screensMethacrylate screens

installed in front of each rack

Benefits observed:– All BladeCenters Rack has an

l t t / 1ºCequal temperature +/- 1ºC– BladeCenter fan speed

reduced


Foto: BSC


Results (including other improvements):

reduction of 10% of power consumption (and CO2)consumption (and CO2)

Marenostrum power consumption: approx 1 2

Image courtesy of UPC

consumption: approx. 1.2 Mwats

aprox 1 100 000 €/year


aprox 1.100.000 €/year

Decreasing DC power consumptionconsumption

How can we decrease DC power consumption without a concomitant decrease in the IT load?

Can we move the Marenostrum in a better location?e.g. more efficient cooling due to climate conditions


Can we put a DC in a better location?location?

I containerized marenostrum and it has traveled with me

HelsinkiPUE = 1,073

Estimated ~3% of

Data courtesy of AST IT Infrastructures (www.AST-global.com)

BarcelonaPUE = 1,113


energy could be saved

What else can we do?

Maximize work done per watt


What else can we do?

Maximize work done per watt

Energy consumption is not only determined by theefficiency of the physical resources

but it is also dependent on the resource management

Intelligent management of resources may lead toi ifi d i f h i bsignificant reduction of the energy consumption by

meeting the performance requirements


Contents





Conclusions


Power aware Resource ManagementManagement

Different layers inDifferent layers in a Data Center?

Data Center Infrastructure

Applications

System

Hardware


Different layers in a Data Center

Who is responsible for managing resources?

y

Who is responsible for managing resources?

Data Center Infrastructure …W b S i

ApplicationsWeb Services

Aplication Server

Web server

SystemJava Virtual Machine

Operating System

HardwareVirtualization layer

…

MIDDLEWARE


MIDDLEWARE

Power/Energy Aware Resource Management at Middleware levelManagement at Middleware level

Challenge:– Shift the paradigm from “time to solution" to

"kWh to solution”

How?:How?:– Introducing intelligent self-management techniques

– Adding energy efficiency parameter to the list of critical operating parameters to control that already including

il bilit li bilit favailability, reliability, performance, …


Contents





Conclusions


Reseach at a Glance

Today the most common techniques/control knobs of the middleware are related to:– consolidation and virtualization technology, – turning on/off policies,

new processors– new processors,– dynamic voltage and frequency scaling, – hybrid architectures on datacenters, y ,– federating clouds– etc.


Workload Consolidation

Rationalize and reduce operational costs: server consolidation– implies combining workloads from separate applications into a

smaller number of systems if there are different peak timessmaller number of systems if there are different peak times

C1

rce

emen

t

C2

rce

emen

t

Time

Res

our

Req

uire

Time

Res

our

Req

uire

C1C2

sour

ce

quire

men

t


Time

Res

Req

Virtualization Technologies

Consolidation uses virtualization to share the resources– Virtualization software divides a physical server into isolated virtual

environments, enabling organizations to run multiple applications or OS on a single serverOS on a single server.

Operating System

App 3

Operating System

App 1 App 2

CPU Mem

Virtual Machine

CPU Mem

Virtual Machine

. . . . . .

Operating System

App 1 App 3App 2

CPU CPU Mem

Physical Machine CPU CPUCPU CPU Mem

Physical Machine CPU CPU

Operating System


The energy savings can be up to 20% - 70%

Using multi-core processors

Middleware allow that applications benefit from new multi-core designs beyond the traditional parallelism

Example: Exploitation of Accelerators in memory-boundExample: Exploitation of Accelerators in memory bound Applications using memory compression techniques

App 3App 1 App 2

Operating System

App 3

Virtual Machine

Operating System

App 1 App 2

Virtual Machine

. . . . . .CPU MemCPU Mem

Mem

CPU CPU Mem

Physical Machine CPU CPUCPU CPU

F b d l th i b t 25%


For memory-bound appl. the energy savings can be up to 25%

Use low-power processors

Low-power processors vs high-performance processors

Good approach for transactional workloads (webs)– Experiments at BSC:

Reduction of 70%


Turn on/off policies

However the numerical applications do not behave like this

Experiments at BSC with 4 HPC tasks:tasks:

• 1 Xeon * 1 hour 317 Watts • 2 Atom * 5 hours 398 Watts

– The most energy-efficient approach– The most energy-efficient approach in this environment is to run jobs very fast and then turn off the power of the system


power of the system.

Hybrid hardwarey

But in general the workload in data centers is heterogeneous

F HPC X k ffi i tl th At– For HPC, Xeons work more efficiently than Atoms

– For Transactional Atoms work more efficiently than XeonsFor Transactional, Atoms work more efficiently than Xeons

Therefore, why , ynot build the two together?toget e


Courtesy of Giovanni Leonardi (Mataró-Sicilia)

Hybrid hardwarey

Better tradeoff in terms of energy and revenue– Experiments at BSC with heterogeneous workload:

Reduction of 25%


In the Cloud Era

Resource Management in a Cloud Federationg– Distribution of workload across geographically

distributed DC

– Consolidation at DC level (and turn off a whole DC)

– Possibility to reallocate the workload to a place where energy or cooling is cheaper

e.g. solar energy during daytime across different time zones, efficient cooling due to climate conditions etcconditions, etc.

– Etc


Etc.

In the Cloud Era

Outsourcing/Insourcing strategiesg g g

The energy savings can be up to 10%? 30? 50%?


Conclusions

The energy savings can be up to 20% - 70%

For memory-bound appl the energy savings can be up to 25%Reduction of 70%Reduction of 25%

For memory-bound appl. the energy savings can be up to 25%The energy savings can be up to 10%? 30? 50%?

Resource management at Middleware level will play animportant role in the reduction of energy consumption

Further research is required!


Conclusions

Future research directions in resource management– Development of energy-efficient resource management approaches for

multi-core CPUs due to the wide adoption of these.

– Development of intelligent techniques to manage resources efficienty.– Efficient workload distribution across geographically distributed DC.

– SLA negotiation between resource providers and users that includescontrol over power consumption and/or CO2 emissions.control over power consumption and/or CO2 emissions.

– Etc.




Middl h Middleware research for green Data Centers

Prof. Jordi TorresBarcelona Supercomputing Center (BSC)UPC Barcelona TechUPC – Barcelona Tech




Jordi Torres has a Master degree in Computer Science from the TechnicalUniversity of Catalonia (UPC Barcelona Tech, 1988) and also holds a Ph. D.

About the author:y ( )

from the same institution (Best UPC Computer Science Thesis Award,1993).Currently he is a full-time professor in the Computer ArchitectureDepartment at UPC and is a manager for the Autonomic Systems andDepartment at UPC and is a manager for the Autonomic Systems andeBusiness Platforms research line at Barcelona Supercomputing Center(BSC).His current principal interest as a researcher is to make IT resources more

ffi i t i d t bt i t i bl IT d t f thefficient in order to obtain more sustainable IT and to focus on the resourcemanagement needs of modern distributed and parallel cloud computingenvironments. He has worked on a number of EU and industrial researchand development projects. He has more than a hundred publications in the

d i l d i l f i th fi ld f di t ib t d darea and was involved in several conferences in the field of distributed andparallel systems.He has been Vice-dean of Institutional Relations at the Computer ScienceSchool, and currently he is a member of the University Senate and member


, y yof the Board of Governors.

He can be reached at www.JordiTorres.org

workshop on green and environmental computing - helsinki

Education