There is a wide variety of tools for the collection and analysis of performance data for parallel applications. However, the use of external tools for performance tuning is too laborious a process for many applications, especially for a complex integrated system such as Rocstar. Many tools are not available on all the platforms for which we wish to collect data, and using a disparate collection of tools introduces complications as well. To obviate the need for external tools, we have extended the Roccom framework's ``service-based'' design philosophy with the development of a performance-profiling module, Rocprof.
Rocprof provides two modes of profiling: module-level profiling and submodule-level profiling. The module-level profiling is fully automatic, embedded in Roccom's function invocation mechanism, and hence requires no user intervention. For submodule-level profiling, Rocprof offers profiling services through the standard MPI_Pcontrol interface, as well as a native interface for non-MPI based codes. By utilizing the MPI_Pcontrol interface, applications developers can collect profiling information for arbitrary, user-defined sections of source code without breaking their stand-alone codes. Internally, Rocprof uses PAPI[4] or HPM (http://www.alphaworks.ibm.com/tech/hpmtoolkit) to collect hardware performance statistics.