Automatic Profiling of MPI Applications with Hardware Performance Counters

摘要: This paper presents an automatic counter instrumentation and profiling module added to the MPI library on Cray T3E SGI Origin2000 systems. A detailed summary of hardware performance counters calls any production program is gathered during execution written in MPI_Finalize a special syslog file. The user can get same information different Statistical summaries are computed weekly monthly. describes experiences with this systems at HLRS Stuttgart TU Dresden. It focuses problems integrating into first results these counters. Also, second software design described that allows integration layer dynamic shared object without consuming user's PMPI interface.