From: Leigh Orf Newsgroups: comp.parallel.mpi,comp.parallel.pvm Subject: Poor MPI performance with collection of Linux workstations Date: 17 Mar 1999 03:35:53 GMT Organization: Berbee Information Networks Corporation Message-Id: <7cn7up$4tq$1@grandprime.binc.net> References: <36ED325A.69D8C9B8@rrsg.ee.uct.ac.za> User-Agent: tin/pre-1.4-981225 ("Volcane") (UNIX) (Linux/2.2.3 (i586)) Xref: ukc comp.parallel.mpi:4762 comp.parallel.pvm:8129 Mark Gebhardt wrote: > Hi all [big snip] > Under certain conditions, my MPI implementation gives wildly varying > execution times - something that the PVM version never did. Here's > what happens: I don't have an answer to your problem, but I have a similar problem. I'm running Linux and have just recently upgraded all machines to the 2.2.3 kernel, hoping my problems would be solved. My problems persist. I have nine machines, one of which is a 4-processor box. They are all 200 MHz PPros. I get wildly varying execution times, too. I am running an atmospheric numerical model in a 3-d torus architecture through an intel switch, 100baseT, Digital DS21140 Tulip ethernet cards (half duuplex) in each machine. Each machine (node) runs the same code, which goes like (1) calculate stuff (2) exchange boundary data between adjacent nodes (3) repeat. I would expect all machines to run at a load level approaching 1.0, but I get load levels that are much lower on each node. One time block will integrage in ~1000 seconds, the next ~3000. This happens for 4x2 and 3x3 geometries. When I run the model only on the quad box, which has 512 MB per node, I get the same kind of poor performance with a fluctuating load level. I am about to try two things: first putting in nupshot code, and also writing a PVM version of my model, which should be pretty straightforward. I figure it must be something with communication between nodes, but I don't know where to look. For a look at what a snapshot of xloads look like during a typical run, see: http://scat.ssec.wisc.edu/orf/~xloads/ Any pointers would of course be appreciated. Leigh Orf Email: orf < at > mailbag < dot > com