Hi All I'm the QA Manager at Allinea, a small company that produces the DDT Parallel Debugger and MAP Parallel Profiler, and we spend a lot of time manipulating MPI environments to get our debugging and profiling to work. In particular for the MAP profiler, which is a sampling one, we want to record time spent in MPI calls and metrics like bytes send/received. Just observing the stack trace proved unsatisfactory, so we now have a wrapper library for MPI routines that disables sampling during MPI calls and collects timing information and metrics manually.
We want to ensure we've wrapped all key routines (ones that involve significant communication and time) and test that our sampling technique is robust. I did have preliminary discussions with Jeff Squyres in the spring about using MTT for testing DDT/MAP, and we may still want to go down that route, but I suspect its rather overkill for what I want at the moment, which is to run MAP against a simple test code that calls all known MPI functions for a few seconds each, and see how our wrappers cope. Does anyone know of a simple test code that exercises all the routines in the MPI 1 and 2 standards that I could use? Thanks John
