On Jan 26, 2010, at 4:22 AM, Mathieu Gontier wrote:
1/ I rebuilt without --enable-dist (more secured indeed) and with explicit --without-openib/--with-openib : behaviors are better. Great.
Excellent. I didn't mention it in my prior email, but our configure behaviors are generally:
--with-<foo>: Build with foo support. If foo support is not available, fail/abort configure
--without-<foo>: Don't even check for foo support; don't build it at all.
neither: check for foo support. If foo support is available, build it. If not available, skip it.
2/ Yes, my PATH and LD_LIBRARY_PATH are correctly set
3/ There certainly were previous installations of OpenMPI on this machine, but not in the same directory; before rebuilt, I correctly uninstall the previous installations (thank you for the tip)
4/ Is there a way to have the list of the plugins in OFED?
Not really. The verbs library (the "native" networking API for Openfabrics) has plugins similar to Open MPI, but they don't have an ompi_info-like command that shows what plugins are available. IIRC, verbs has only 1 type a plugin: to service an underlying device type. IIRC, the names of the drivers are:
- nes: NetEffect / Intel RNICs
- cxgb3: Chelsio T3 RNICs
- mlx: Mellanox ConnectX HCAs
- mthca: Mellanox earlier HCAs
- ipath: QLogic HCAs
That's off the top of my head and it may not be a comprehensive list. Check the installed software on your machine to see what you have.
5/ Yes, good messages about the device will be welcome, but with 1/ it is really better now
6/ The message is really more explicit explained like that (thanks)
7/ I built both my small test and OpenMPI-1.4.1 directly on my cluster with gcc-3.4 and I still have this error. Do you have any idea where the problem could come from?
--> Returned value -13 instead of OPAL_SUCCESS
That's an odd one. Can you send your exact configure line again, and perhaps the output from configure? (please compress -- see http://www.open-mpi.org/community/help/).