Hi, We (Debichem) also have a few upstreams that claim OpenMPI is more buggy than MPICH.
Of course, that mustn't be the case by the time buster is released. On Fri, Jul 07, 2017 at 09:22:52AM +0100, Alastair McKinstry wrote: > The bigger question here is : how do we decide between them? How about we have both, at least for the core linear algebra packages, like scalapack etc.? > At $work when we make the decision on default MPI to recommend when we > install a new cluster, testing our standard codes and benchmarking: > first does it compile and work (nearly always), secondly whats the speed > / latency / variance? > > Given our user base, the most useful criteria is probably: does it have > the necessary feature base. We can't benchmark as we don't know general > users' hardware. OpenMPI appears to have more complete hardware > support, and should be the default unless it is shown to be 'too buggy'. > > How to test ? can we go through our package base and turn on MPI > threading in testing? There's http://packages.debian.org/src:mpi-testsuite which is basically the MPICH testsuite, but OpenMPI runs on it as well. The idea was to see how the various implementation work on the various architectures. I haven't worked on that package in a while and/or analysed the results, so if somebody want to help/take it from here, be my guest. Michael
