OpenMP 3.0 Profiling: a paper and an implementation
Oleg and I wrote a paper describing the techniques used to do the OpenMP 3.0 performance profiling which was mentioned in my previous blog. The paper is going to be presented at this year’s IWOMP.
In the meantime, Sun has just released Sun Studio Express 3/09.
A new collector for OpenMP 3 profiling is available, but not enabled by default. It can be enabled by setting the environment variable SP_COLLECTOR_NEWOMP, and works only for code compiled with the option -qoption iropt -Apcg:mfcxt. If you enable the new collector, the following features are available:
* The Analyzer GUI includes OMP Parallel Regions tab and OpenMP Tasks tab, which are available when OpenMP data is collected, and have entries for the source lines of each construct and metrics for those entries.
* Support has been added in the experiment format for OpenMP 3.0 profiling.
* Two new commands for OpenMP profiles, OMP_preg and OMP_task have been added.
* User mode presentation of OpenMP profiles has been changed so that the parallel loop functions are no longer shown.
