CERN openlab II - Platform CC - Optimization

Benchmarking

Benchmarking, or measuring the performance, has always been the best way to describe the capabilities of a certain CPU or computer. The main challenge however, is to find a benchmark which reflects the actual workload reasonably well. Up to now the workload generated by the CERN/HEP experiments was scaling quite well with the performance measured with the SPECint2000 benchmark. But already with SPEC certain difficulties appear. Depending on the compiler and the compiler options the results differ significantly.

In order to obtain useful results CERN uses a specific configuration for running the SPEC2000 benchmark.
The benchmarks are compiled with gcc and the options "-O2 -fPIC -pthread".

• -O2 - standard optimization flag

• -fPIC - produces "Position Independent Code", used mainly for shared libraries

• -pthread - try to use POSIX threads

The benchmarking project inside openlab has several purposes:

• better understanding of CERN workload (in connection with the performance monitoring project)

• evaluating the new SPEC2006 benchmark

• look at possible correlations with CERN/HEP workload

• possible replacement of SPECint2000 as standard benchmark

• test and evaluate pre-production systems provided by Intel

For more information on the frameworks for LHC experiments and the compiler project, see the website edited by Jose Dana. A benchmark repository, edited by Havard Bjerke is also available to you.