FCA - problem with MPI_Allgather(v) when sendbuf is MPI_IN_PLACE

[ompiteam]

MPI standard says that if MPI_Allgather() argument sendbuf is MPI_INPLACE the arguments sendcount and sendtype should be ignored

Look at the enclosed test case. It passed if FCA Allgather is disabled

[aryzhikh@srvmpidev03 bug_fca_allgather]$ mpirun -n 2 -npernode 1  -host srvmpidev03,srvmpidev04   -x LD_LIBRARY_PATH ./test
Test passed
Test passed

And the test fails when FCA Algather is enabled:

[aryzhikh@srvmpidev03 bug_fca_allgather]$ mpirun -n 2 -npernode 1  -
host srvmpidev03,srvmpidev04  -mca coll_fca_np 0 -mca coll_fca_enable 1  -mca coll_fca_priority 10000 -mca coll_fca_verbose 0 -x LD_LIBRARY_PATH ./test

[0] buf[1]=-1 but expected 1
Test failed
[1] buf[0]=-1 but expected 0
Test failed

I tried Open MPI 1.6.2 with FCA 2.1
The failure happens because of the value of sendcount (-1)

Proposed fix of given bug is attached

[ompiteam]

Imported from trac issue 3342. Created by aryzhikh on 2012-10-04T06:21:08, last modified: 2013-12-13T13:48:31

• aryzhikh attached test.tar.gz on 2012-10-04 06:22:18
• aryzhikh attached fix.tar.gz on 2012-10-04 06:23:32

[ompiteam]

Trac comment by jsquyres on 2013-12-13 13:48:31:

Mellanox - should this be committed to the v1.6 branch?

[mike-dubman]

@vasilyMellanox please take a look and fix.
Thanks

[jsquyres]

@miked-mellanox FYI: note that we decided this past Tuesday that 1.6.6 will not happen. All focus is on the v1.8 series and upcoming v1.9 series.