Hi Scott,
What is happening is that on creation of the Queue Pair the max inline
data is reported as 0. Open MPI 1.0.1 did not check this and assumed
that data smaller than some threshold could be sent inline :-(. The
Open MPI trunk does check the max inline data QP attribute prior to
using inline data for sending, but if it sees that the inline data is 0
it will report a warning although the application should still run to
completion. Is your application completing with the trunk snapshot, if
so you can effectively ignore these warnings?
We saw this problem (0 max inline data) on one of our clusters,
upgrading the OpenIB stack corrected the issue. You are running a 2.6.9
kernel however so to upgrade to the latest OpenIB stack you will need
to get a backport which is in the OpenIB SVN under
gen2/branches/backport-to-2.6.9.
So to distill, the hang on 0 max inline data was corrected in Open MPI
in the trunk and the 1.0.2 branch. Upgrading the OpenIB stack resulted
in a positive max inline data, although I cannot explain why, there may
have been other changes to the cluster I am not aware of, Roland could
probably explain this better.
- Galen
On Feb 28, 2006, at 1:36 PM, Scott Weitzenkamp ((sweitzen)) wrote:
I'm trying OpenMPI with the 2.6.9-27.ELsmp kernel and Red Hat OpenIB
RPMs included with the beta.
I ran configure with only the --prefix option.
OpenMPI 1.0.1 mpirun just hangs with no output.
A 1.1 snapshot gets this:
/data/software/qa/MPI/openmpi-1.1a1r9094.rhel4-amd64-openib/bin/mpirun
--prefix
/data/software/qa/MPI/openmpi-1.1a1r9094.rhel4-amd64-openib -np 2
-host 192.168.0.197,192.168.0.199
/data/home/releng/tmp/rhel4/openib/mpi_latency.amd64 100 1
[0,1,1][btl_openib_endpoint.c:889:mca_btl_openib_endpoint_create_qp]
ibv_create_qp: returned 0 byte(s) for max inline data
[0,1,0][btl_openib_endpoint.c:889:mca_btl_openib_endpoint_create_qp]
ibv_create_qp: returned 0 byte(s) for max inline data[
0,1,1][btl_openib_endpoint.c:889:mca_btl_openib_endpoint_create_qp]
ibv_create_qp: returned 0 byte(s) for max inline data
[0,1,0][btl_openib_endpoint.c:889:mca_btl_openib_endpoint_create_qp]
ibv_create_qp: returned 0 byte(s) for max inline data
[0,1,0][btl_openib_endpoint.c:889:mca_btl_openib_endpoint_create_qp]
ibv_create_qp: returned 0 byte(s) for max inline data
[0,1,0][btl_openib_endpoint.c:889:mca_btl_openib_endpoint_create_qp]
ibv_create_qp: returned 0 byte(s) for max inline data
[0,1,1][btl_openib_endpoint.c:889:mca_btl_openib_endpoint_create_qp]
ibv_create_qp: returned 0 byte(s) for max inline data
[0,1,1][btl_openib_endpoint.c:889:mca_btl_openib_endpoint_create_qp]
ibv_create_qp: returned 0 byte(s) for max inline data
Do I need to use a newer OpenIB?
Scott
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users