[Mvapich-discuss] MVAPICH2 GDR from source code?

[John Moore]

Hello,

We have been struggling to get MVAPICH2 to work with cuda-aware support and
RDMA. We have compiled MVAPICH2 from source, with the --enable-cuda option,
but when we run the osu_bibw bandwidth test using Device to Device
communication, we get a segmentation fault.

Below is the output from osu_bibw using MVAPICH2:
 MVAPICH2-2.3.6 Parameters
---------------------------------------------------------------------
        PROCESSOR ARCH NAME            : MV2_ARCH_AMD_EPYC_7401_48
        PROCESSOR FAMILY NAME          : MV2_CPU_FAMILY_AMD
        PROCESSOR MODEL NUMBER         : 1
        HCA NAME                       : MV2_HCA_MLX_CX_HDR
        HETEROGENEOUS HCA              : NO
        MV2_EAGERSIZE_1SC              : 0
        MV2_SMP_EAGERSIZE              : 16385
        MV2_SMP_QUEUE_LENGTH           : 65536
        MV2_SMP_NUM_SEND_BUFFER        : 16
        MV2_SMP_BATCH_SIZE             : 8
        Tuning Table:                  : MV2_ARCH_AMD_EPYC_7401_48
MV2_HCA_MLX_CX_HDR
---------------------------------------------------------------------
# OSU MPI-CUDA Bi-Directional Bandwidth Test v5.7.1
# Send Buffer on DEVICE (D) and Receive Buffer on DEVICE (D)
# Size      Bandwidth (MB/s)
1                       0.07
2                       0.15
4                       0.29
8                       0.57
16                      1.12
32                      2.30
64                      4.75
128                     9.41
256                    18.44
512                    37.22
1024                   74.82
2048                  144.70
4096                  289.96
8192                  577.33
[cell3:mpi_rank_0][error_sighandler] Caught error: Segmentation fault
(signal 11)
[cell3:mpi_rank_1][error_sighandler] Caught error: Segmentation fault
(signal 11)

===================================================================================
=   BAD TERMINATION OF ONE OF YOUR APPLICATION PROCESSES
=   PID 471850 RUNNING AT cell3
=   EXIT CODE: 139
=   CLEANING UP REMAINING PROCESSES
=   YOU CAN IGNORE THE BELOW CLEANUP MESSAGES
===================================================================================
And this is with OpenMPI:
# OSU MPI-CUDA Bi-Directional Bandwidth Test v5.8
# Send Buffer on DEVICE (D) and Receive Buffer on DEVICE (D)
# Size      Bandwidth (MB/s)
1                       0.43
2                       0.83
4                       1.68
8                       3.37
16                      6.72
32                     13.42
64                     27.02
128                    53.78
256                   107.88
512                   219.45
1024                  437.81
2048                  875.12
4096                 1747.23
8192                 3528.97
16384                7015.15
32768               13973.59
65536               27702.68
131072              51877.67
262144              94556.99
524288             157755.18
1048576            236772.67
2097152            333635.13
4194304            408865.93


Can GDR support be obtained by compiling from source like we are trying to
do or do we have to use an RPM? We export MV2_USE_CUDA=1. Any
recommendations would be greatly appreciated.

Thanks,
John
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osu.edu/pipermail/mvapich-discuss/attachments/20220111/aa671d3a/attachment-0021.html>