Hello all,
I had run OpenMC using 2 similar nodes. and the timings are as expected.
Now I have more nodes with a different configuration (different number of sockets/node, core/socket, threads/core).
Can somebody help me in running and distributing job with MPI in this type of cluster configs?
Thank you.
Hi Sachin,
Unfortunately at present there’s no good way to run across a heterogeneous system like this. Right now, OpenMC divides the number of particles per batch over the number of MPI ranks evenly, so if one MPI rank resides on a node that is faster than another, it will reach the end of the batch sooner and then will have to wait for the other ranks to finish. The only suggestion I have is to run independent jobs across each node type so that there is no load balancing issues within a single job, and then combine the results from the independent jobs.
Best regards,
Paul
Thank you Paul for the suggestions…