From: Alessandro Cembran (cembran_at_chem.umn.edu)
Date: Tue Jun 13 2006 - 17:40:59 CDT
By looking at the ApoA1 performance benchmarks it seems that NAMD should
scale very well up to 64 processors on Altix machines.
By using the pre-compiled Linux-ia64-MPT with a ~190,000 atoms system,
PME, 24 StepsPerCycle and 3 PairlistsPerCycle I get good scaling with 8
processors but a very bad scaling with 32 processors. That is, in going
from 1 to 8 to 32 processors the relative speedups are 1, ~7.5 and ~10.0.
I am also setting the following variables in the submission command:
export MPI_DSM_DISTRIBUTE=yes [ as indicated on pg 114 of the manual ]
export MPI_MEMMAP_OFF=1 [ to avoid the "running out of
memory" problem when submitted to the queue system ]
I was wondering what is the way to achieve the claimed performances,
whether it is necessary to recompile locally namd or to set other
environment variables. I'm aware that increasing the StepsPerCycle in
order to reduce the message-passing should improve the performance, but
I don't think that will help to recover by a factor of three like it
seems the case here for 32 processors.
Thanks in advance,
-- Alessandro Cembran,PhD Post Doctoral Associate Univ. of Minnesota, Dept. of Chemistry G2, 139 Smith Hall 207 Pleasant St SE Minneapolis, MN 55455-0431 Office: +1 612-625-4319 cembran_at_chem.umn.edu
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:43:42 CST