[gmx-users] request for patch and test program from 'Speeding up Parallel GROMACS'
ckutzne at gwdg.de
Wed Apr 4 09:53:50 CEST 2007
David Konerding wrote:
> I was happy to see the 'Speeding up Parallel GROMACS' paper. I'd like
> to try
> the GROMACS patch for better all-to-all, as well as the all-to-all test
> program referred
> to on the paper. I can't find it on the GROMACS site.
you can download both at
www.gromacs.org -> download -> user contributions -> contributed software
If you sort by name, it's on page 2, called gmx_alltoall. The all-to-all
test program is also included in the tarfile. Packet drops during all-to-all
communication will lead to very different individual all-to-all timings,
which is easily seen in the output.
> We've got a Cisco 6500 and when I run tcpdump during MPI runs I see a
> huge amount
> of TCP lost packets with retransmits, and tcptrace confirms there are a
> large number of restransmits.
> I am trying to get our network admins to test the flow control setting
> option on our 16-processor cluster.
From my experience I can say that on 16 CPUs you will need either flow
control or an ordered all-to-all to avoid packet loss. If you can enable
flow control, you should definitely do that, because in principle there
can be congestion in many parts of a program, the (lam-) all-to-all just
turned out to be the No. 1 candidate.
Dr. Carsten Kutzner
Max Planck Institute for Biophysical Chemistry
Theoretical and Computational Biophysics Department
Am Fassberg 11
37077 Goettingen, Germany
Tel. +49-551-2012313, Fax: +49-551-2012302
More information about the gromacs.org_gmx-users