[gmx-users] Re: Gromacs in parallel
David van der Spoel
spoel at xray.bmc.uu.se
Tue Jul 23 09:46:10 CEST 2002
On Mon, 22 Jul 2002, Brian Smith wrote:
>Some time ago I enquired about running GROMACS in parallel; we had a
>problem where running over 3 nodes took three time longer to complete).
>You had correctly identified that I hadn't initalised LAM correctly (ie.
>We have progressed from this situation to another, more puzzling one.
>We run a simulation on 2 and 4 nodes. The MEGA-FLOPS accounting suggests
>we are still NOT seeing improvements with parallelisation,...
>2-nodes: NODE (s) =103.000, Real(s)=103.000
>4-nodes: NODE(s)=142.000, Real(s)=142.000
>When I inspect the nodes during the run, each node is runing at 1/N of the
>CPU, where N=number of nodes (ie. with 2 nodes, each nodes is running 50%
>CPU use, and with 4 nodes, each node runs at 25% CPU use).
>Can you suggest what we are doing wrong?
You have to be more specific: how was LAM compiled, what kind of nodes are
these (dual/single processor) and what processor. E.g. the intel P4 xeon
has a feature called hyperthreading that you have to turn off...
>Many thanks for any hints you can give.
Dr. David van der Spoel, Biomedical center, Dept. of Biochemistry
Husargatan 3, Box 576, 75123 Uppsala, Sweden
phone: 46 18 471 4205 fax: 46 18 511 755
spoel at xray.bmc.uu.se spoel at gromacs.org http://zorn.bmc.uu.se/~spoel
More information about the gromacs.org_gmx-users