From: Floris Buelens (floris_buelens_at_yahoo.com)
Date: Sat Jun 18 2005 - 08:25:39 CDT
I'm running NAMD on a cluster of dual Xeon systems.
For my purposes I need to run a large number of
non-interdependent MD simulations, (starting from the
same point with slightly different parameters), so the
way to go is clearly to run each only on one or two
CPUs at once so there's no communication overhead. As
most of the discussion out there seems to be about
improving scaling for large clusters, I'd be grateful
for any comments that might help me get the most out
of the setup.
Currently I'm using the pre-compiled release binaries
to run "charmrun namd2 ++local +p4" - 4 threads on the
2 CPUs gives an improvement over 2, I guess through
the magic of hyperthreading.
My questions are:
- How are these 4 internal threads communicating when
I use the precompiled binaries? Is there a way that
might be more efficient given that inter-process
communication should be extremely fast?
- Is there anything to gain from recompiling NAMD,
with optimisations either for Xeon (3ghz, SSE2, HT),
or for this kind of 'slightly parallel' setup? Or for
a long shot, does anyone have binaries like that lying
Thanks in advance!
Birkbeck College, UK
Yahoo! Messenger - NEW crystal clear PC to PC calling worldwide with voicemail http://uk.messenger.yahoo.com
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:39:35 CST