CCL: Changing SYSV parameters for parallel GAMESS
- From: Dr Seth OLSEN <s.olsen1~at~uq.edu.au>
- Subject: CCL: Changing SYSV parameters for parallel GAMESS
- Date: Wed, 10 Nov 2004 17:41:35 +1000
Dear Collegues,
I recently installed and ran parallel GAMESS using DDI on a 2CPU Xeon machine
with hyperthreading (4 virtual CPU's). I ran through the entire included test
set using 4, 2 and 1 cpus and found interesting results for the timings. The
time to run gamess seemed to be by far lowest with 1cpu, and 2,4 cpus scaled
about linearly against each other. However, running on 2 or 4 nodes resulted in
very, very low cpu usage (<10%) and longer cpu and wall times. At this point
I extended the sysV shared memory from 33MB to 1GB, thinking that this was why
there was no improvement in time with parallel use. What I found after this was
that 1) the computer was significantly slowed down for everything, even mundane
tasks 2) the cpu usage was marginally improved for 2 & 4 node use, but was
reduced for 1 node use 3) the wall time was made longer in all cases and had no
obvious correlation with the number of nodes used.
Can any of the discussion group member shed light on these observations? Should
I have increased the SYSV shared memory? Should I increase the number of
semaphores as well? What kind of improvements should I expect using parallel
GAMESS?
Cheers,
Seth Olsen
ccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccms
Dr Seth Olsen, PhD
Postdoctoral Fellow, Computational Systems Biology Group
Centre for Computational Molecular Science
Chemistry Building,
The University of Queensland
Qld 4072, Brisbane, Australia
tel (617) 33653732
fax (617) 33654623
email: s.olsen1~at~uq.edu.au
Web: www.ccms.uq.edu.au
ccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccms