Speaker
Dr
Gabriele Benelli
(CERN PH Dept (for the CMS collaboration))
Description
The demanding computing needs of the CMS experiment require thoughtful planning and management of its computing infrastructure. A key factor in this process is the use of realistic benchmarks when assessing the computing power of the different architectures available. In recent years a discrepancy has been observed between the cpu performance estimates given by the reference benchmark for HEP computing (SPEC INT) and actual performances of HEP code. Making use of the cpu performance tools from the CMSSW performance suite, comparative cpu performance studies have been carried out on several architectures. A benchmarking suite has been developed and integrated in the CMSSW framework, to allow computing centers and interested third parties to benchmark architectures directly with CMSSW. The CMSSW benchmarking suite can be used out of the box, to test and compare several machines in terms of CPU performance and report with the wanted level of detail the different benchmarking scores (e.g. by processing step) and results. In this talk we describe briefly the CMSSW software performance suite, and in detail the CMSSW benchmarking suite client/server design, the performance data analysis and the choice and composition of the benchmark scores. The interesting issues encountered in the use of HEP code for benchmarking will be discussed and CMSSW benchmark results presented.
Primary author
Dr
Gabriele Benelli
(CERN PH Dept (for the CMS collaboration))
