On Fri, 2017-07-07 at 09:22 +0100, Alastair McKinstry wrote: > Yes, I agree. > > The bigger question here is : how do we decide between them? .. > How to test ? can we go through our package base and turn on MPI > threading in testing?
Some testsuites are already at hand, e.g. scalapack-mpi-test. My feeling is there might be limited value in running these simple tests for benchmarking purposes, since they aim at confirming the code runs, rather than checking performance of how well it runs. It could expose any plain threading bugs. There is mpi-testsuite, pulled from mpich but applied to openmpi. Is it sufficient for this kind of testing or would we need longer, tougher benchmarks to justify any switch in mpi-defaults? Given the hardware advantages of openmpi, we'd need benchmarks to show a clear-cut advantage in mpich. That would mean winning over a range of different benchmarks (different applications), not just one test. Drew
