Bryan Lally wrote:
Here's what we've found. It wasn't the platform file as such. I've
since built with ./configure and some standard, obvious command line
switches. What's then required is to edit the platform configuration
file, <prefix>/etc/openmpi-mca-params.conf and add:
coll_sync_priority = 100
coll_sync_barrier_before = 1000
Oops. Hit send a bit before I was ready.
This has eliminated the problem on two Fedora 9 machines (8 cores and a
2 core laptop) and a 4 core Fedora 7 machine.
Thanks to all who helped get this figured out, particularly Ralph.
- Bryan
--
Bryan Lally, la...@lanl.gov
505.667.9954
CCS-2
Los Alamos National Laboratory
Los Alamos, New Mexico
