News 

Events

Press Corner

Press Releases

Press Coverage

Multimedia Corner

Partners' Spotlights

Documents

Annual Reports

Newsletter

Technical Documents

Presentations

 

openlab Phase III

Automation Controls CC

Database CC

Networking CC

Platform CC

Previous Phases

Management

Education Corner

 

Student Programme

What is it?

How to apply-2012

Students-2012

Programme-2012

About CERN openlab

What is it?

Participants

Guiding Principles

 

openlab II

Platform CC

Virtualization

Optimization

Grid IC

Tycoon

SmartDomains

Database CC

Networking & Security

Networking
Security

openlab I

opencluster

Usage

LCG on Itanium 2

SmartFrog

 

CERN openlab II - Platform CC - Optimization

Thermal Testing

The CERN computing center has a limitiation for electric power of 2.5MW. Given this hard limit, we are naturally interested in getting the most compute power out of this electric power.

To be useful the thermal testing of a server has to go hand in hand with testing the performance of this server. One of the main metric by which a servers capabilities can be judged is "performance per Watt" which is "SPECint2000/Watt" in our case.

But the project goes beyond just measuring power consumption under load. We also try to better understand where in a server the power is consumed. The CPUs have been the main power consumer for a long time now, but with the new generation of very power efficient CPUs, the power consumption of the main memory comes into focus now.

Some of the power measurements results so far:

  • FB-DIMM (memory in Woodcrest servers): ~10W per GByte under full load

  • with 2GB per core this means 80W for the memory alone in a Woodcrest

  • hard disks: ~10W idle and ~15W under load (of course, only the models we tested, other models may have different power consumptions)

As for SPECint2000/Watt measured with CERN settings the results look as follows:
(the results are only meaningful for a specific system (type) in connection with the used compiler, and possibly additional libraries!)

  • 2.6GHz Nocona (2GByte RAM): <9 SPECint2000/Watt

  • 2.66GHz Woodcrest (8GByte RAM): ~20 SPECint2000/Watt

Those numbers depend on a lot of factors. Most prominently the used compiler.

If the Intel compiler is used with the highest optimization, the performance is much better: ~30 SPECint2000/Watt.