SMP performance on Scyld
mcosta at fc.up.pt
Mon Dec 3 04:34:38 PST 2001
Wonder if you can help me again.
I'm testing Scyld on a cluster of dual athlon workstations and when I
spawn MPI processes on all nodes there's always one process per node
that is a lot faster than the other, so that most of the time only one
process is running (this doesn't happen on RedHat 7.2 + LAM).
it's funny because running a single process is faster that with Redhat.
Isn't beompi configured to use shared memory on smp's, could this be
because of bad smp support on 2.2.x kernels, or what?
Using the mpirun flags -arch smp etc. doesn't change a thing.
I'm guessing this must be a common problem, with a simple solution...
More information about the Beowulf