[gmx-users] Unable to reproduce DPPC benchmark

[Pietro Lopriore]

Dear all,
we have some CPU-loading problems with a Power970 Linux Cluster with LAM
and Myrinet, all compiled with gcc.
While running the dppc benchmark we're unable to reproduce the same timing
results from time to time.
Checking the processes running on the different nodes, sometimes happens
that some processors are not fully loaded while others are at 99.9%.
This behaviour seems quite strange because it is not reproducible and I
don't think it's connected to a single node or processor (that sometimes
work fine and others not).
So sometimes we have the DPPC (with 200 steps) on 16 CPUs finished in 23
seconds while others in more than 110 seconds. Is this behaviour normal? Is
it GROMACS that doesn't balance the load to the various processors or do
you think that could be something else (like Myrinet drivers, etc.)?
I ensure you that the cluster is empty and that there are no "ghost"
processes on the nodes while running GROMACS.
Any ideas?
Many thanks for your time.

Pietro