With the fix from Johannes to the build script, I'm now down to 38 seconds (from 52) using the HashDist build, while the Ubuntu 1.3 packages gives 40 seconds, so I'm very happy for now. Thanks everyone for helping out in figuring this out.
-- Anders ons 1 apr. 2015 kl 11:46 skrev Jack HALE <[email protected]>: > I guess that depends on what resources you request and how the cluster > is setup! PETSc devs recommend binding MPI processes to sockets too, > so I presume that applies to clusters as well. > > I think vader backend will be useful regardless of whether your > machine is in a cluster or not. My understanding is that OpenMPI knows > about the hierarchy of interconnects and will use the fastest one > available. So you can use the btl ordering self,vader,openib on a > cluster and local processes will use the vader mechanism to transfer > messages whilst remote processes will still use infiniband. The > standard hierarchy for btl is something like self,sm,openib. vader > with a cma backend will always beat sm, so if you can you should use > it. > > Next year we will get another machine so we will have two large NUMA > machines with infiniband. Then I can give some more definite answers > to your questions. > ----- > Dr. Jack S. Hale > University of Luxembourg > _______________________________________________ > fenics mailing list > [email protected] > http://fenicsproject.org/mailman/listinfo/fenics >
_______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
