CPU clock frequency is not likely to be increased significantly in the coming years, and data analysis speed can be improved by using more processors or buying new machines, only if one is willing to change the paradigm to a parallel one. Therefore, performance monitoring procedures and tools are needed to help programmers to optimize existing software running on current and future hardware. Low level information from hardware performance counters is vital to spot specific performance problems slowing program execution. HEP software is often huge and complex, and existing tools are unable to give results with the required granularity. We will report on the approach we have chose to solve this problem that involves decomposing the application into parts and monitoring each of them separately. Both counting and sampling methods are used to allow an analysis with the required custom granularity: from global level, up to the function level. A set of tools (based on perfmon2 – a software interface to hardware counters) for CMSSW, Gaudi and Geant4 has been developed and deployed. We will show how this type of analysis has been proven useful in spotting specific performance problems and effective in helping with code optimization.
展开▼
