It would be best if an IB vendor replies (hint hint!), but it is likely that you have some kind of hardware issue on that node (e.g., a bad / flakey HCA, etc.). You should probably run a full set of layer-0 diagnostics on your fabric to make sure it's clean.
I say this because back when Cisco was an IB vendor, when I ran into weird-o problems like this, they were almost always due to hardware issues (e.g., replace the HCA and then all was fine). Consult your IB vendor's documentation on how to run layer-0 diagnostics. On Nov 4, 2010, at 7:39 PM, Ondrej Marsalek wrote: > Dear all, > > I would like to ask for help with understanding an error message I get > when communication using Open MPI 1.4.1 over Infiniband fails. After > several hours of operation, communication with one particular node > (f24) fails with something like: > > [[20265,1],79][btl_openib_component.c:2951:handle_wc] from f05 to: f24 > error polling LP CQ with status INVALID REQUEST ERROR status number 9 > for wr_id 309134592 opcode 1 vendor error 138 qp_idx 2 > [[20265,1],39][btl_openib_component.c:2951:handle_wc] from f24 to: f05 > error polling LP CQ with status WORK REQUEST FLUSHED ERROR status > number 5 for wr_id 313731584 opcode 1 vendor error 249 qp_idx 2 > > This is reproducible in the sense that it happens repeatedly, but so > far I was not able to create a test case that would trigger the > problem. It happens after hours of smooth operation. One of the nodes > involved is always f24. When I leave it out of the job, I get stable a > run with no trouble. Is this a hardware error or something else? Is > there something I can do try to locate the problem better? Where can I > find what the error codes mean? > > Thanks, > Ondrej Marsalek > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/