Hi Piotr,
A user has reported problems running MPI jobs at several sites while his
job ran perfectly fine at other sites. I ran MPI jobs as well at the
culprit sites and got the same errors. I suspect it has to do with the
fact the home directories of the poolaccounts should be shared between
CE and WNs for MPI jobs but aren't at those sites. (Although these sites
advertise that they can run MPI jobs)
Would it be an idea to include a simple MPI job in your test suite
especially for sites claiming to be able to run MPI jobs. Such a job
should be able to run on something like 4 CPUs to catch badly-configured
sites having dual-CPU nodes.
Best regards,
Ron
|