On Thu, May 26, 2011, Daniel Aiello wrote:
> Therefore, we decided to run the same simulation using sander.MPI with
> 16 processors. Below is the script I submitted to the SGE cluster to do
> this. For comparison purposes, I am also running the serial version,
> using a single processor. The processes for the sander.MPI job are
> being distributed across 4nodes, each with 4 cores. Although, the
> parallel job appears to be running faster than the serial job, the MPI
> simulation runs for much longer than the estimated 5hr runtime, more
> than a day, and is prone to hang ups. We are suspicious that the longer
A better way to benchmark simulations is to run short simulations, since run
time is just proportional to the length of simulation. In this way, you can
quickly see how serial and parallel performance depends on number of
processors, nodes, etc.
Second, if you are looking at performance, using pmemd or pmemd.MPI should
give much better behavior than sander.MPI. Is there a reason for you to be
using sander?
....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu May 26 2011 - 10:00:02 PDT
