Slide 3 of 24

Profiler FeaturesCpu time/HW counters

• Low overhead: 1-10% overhead depending on features

• Precise: not based on sampling

• Function-level and basic-block-level profiling (R2.7)

  -Mprof=func for function level profiling
  
  -Mprof=lines profiles at the 'basic block' level

• CPU time and hardware (HW) counters. (R2.7)

  • Always get cpu time for function level-profiling. Elapsed time and self-time (elapsed minus profiled children).

• Optionally specify 2 HW counters to monitor at function level.(R2.7)

  setenv PROFILE_COUNTERS PP_FLOPS,PP_RESOURCE_STALLS
  

• Get 'self' and elapsed counter values by routine. (R2.7)

• Optionally specify that one of the counters be used for line level profiling instead of using cpu time. (R2.7)

  setenv PROFILE_COUNTERS PP_FLOPS,LINE,PP_RESOURCE_STALLS
  

• List of HW counters with descriptions is on the web:

  • http://www.sandia.gov/ASCI/Red/usage/perfeva.htm

Acknowledgement and Disclaimer