8 messages in org.freebsd.freebsd-clusterLAM MPI on dual processor opteron box...| From | Sent On | Attachments |
|---|---|---|
| Jeffrey Racine | Apr 10, 2004 3:21 pm | |
| Roland Wells | Apr 11, 2004 12:20 pm | |
| Jeffrey Racine | Apr 12, 2004 6:04 am | |
| Bob Willcox | Apr 12, 2004 6:13 am | |
| Jeff Racine | Apr 12, 2004 6:37 am | |
| David O'Brien | Apr 19, 2004 8:32 pm | |
| Jeffrey Racine | Apr 21, 2004 4:41 am | |
| Jeffrey Racine | Apr 21, 2004 4:53 am |
| Subject: | LAM MPI on dual processor opteron box sees only one cpu... | |
|---|---|---|
| From: | Jeffrey Racine (jrac...@maxwell.syr.edu) | |
| Date: | Apr 21, 2004 4:41:52 am | |
| List: | org.freebsd.freebsd-cluster | |
Hi David.
Thanks for your response. With the 4BSD scheduler, things run as expected... lam with 2 processors always fires up cpu 0 and 1 and distributes the load evenly...
PID USERNAME PRI NICE SIZE RES STATE C TIME WCPU CPU COMMAND 838 jracine 101 0 5632K 2012K CPU0 0 0:02 58.94% 5.62% n_lam 839 jracine 101 0 5616K 1968K RUN 1 0:02 57.91% 5.52% n_lam
-- Jeff
On Mon, 2004-04-19 at 23:32, David O'Brien wrote:
On Mon, Apr 12, 2004 at 09:04:24AM -0400, Jeffrey Racine wrote:
Hi Roland.
I do get CPU #1 launched. This is not the problem.
The problem appears to be with the way that current is scheduling.
With mpirun np 2 I get the job running on CPU 0 (two instances on one proc). However, it turns out that with np 4 I get the job running on CPU 0 and 1 though with 4 instances (and associated overhead). Here is top for np 4... notice that in the C column it is using both procs.
PID USERNAME PRI NICE SIZE RES STATE C TIME WCPU CPU COMMAND 96090 jracine 131 0 7148K 2172K CPU1 1 0:19 44.53% 44.53% n_lam 96088 jracine 125 0 7148K 2172K RUN 0 0:18 43.75% 43.75% n_lam 96089 jracine 136 0 7148K 2172K RUN 1 0:19 42.19% 42.19% n_lam 96087 jracine 135 0 7188K 2248K RUN 0 0:19 41.41% 41.41% n_lam
One run (once when I rebooted lam) did allocate the job correctly with np 2, but this is not in general the case. On other systems I use, however, they correctly farm out np 2 to CPU 0 and 1...
Thanks, and any suggestions welcome.
1. Please don't top-post -- it looses context. This is a Unix list, not Mikeysoft one.
2. Have you tried with the 4.4BSD scheduler vs. the "ULE" scheduler? To test, replace: options SCHED_ULE # ULE scheduler with options SCHED_4BSD #4BSD scheduler
-- David