[mvapich-discuss] Segmentation fault at some MPI functions after MPI_Put
khaled hamidouche
khaledhamidouche at gmail.com
Thu Nov 5 10:22:59 EST 2015
Dear, Akihiro,
There seems to be something missing in your setup, as we are not able to
reproduce the segfault locally. As I mentioned your test passes locally.
Further using the Flush we see expected numbers even with IPC=1 (Please see
below). Note that OMB gives the possibility to test all
synchronization/window creation configuration you want to evaluate.
Thanks
../install/bin/mpirun_rsh -np 2 ivy1 ivy1 MV2_USE_CUDA=1 MV2_CUDA_IPC=1
MV2_USE_GPUDIRECT_GDRCOPY=0 ./osu_put_latency -w create -s flush D D
# OSU MPI_Put-CUDA Latency Test
# Window creation: MPI_Win_create
# Synchronization: MPI_Win_flush
# Rank 0 Memory on DEVICE (D) and Rank 1 Memory on DEVICE (D)
# Size Latency (us)
0 0.11
1 14.95
2 14.99
4 14.92
8 14.95
16 14.92
32 14.89
64 14.93
128 14.98
256 14.91
512 14.91
1024 14.99
2048 15.05
4096 15.03
8192 14.96
16384 14.21
32768 15.78
65536 15.58
131072 15.77
262144 16.58
524288 18.74
1048576 27.03
2097152 38.37
4194304 59.76
On Thu, Nov 5, 2015 at 1:14 AM, Akihiro Tabuchi <
tabuchi at hpcs.cs.tsukuba.ac.jp> wrote:
> Dear Khaled,
>
> Thank you for your correspondence.
>
> I inserted a MPI_Win_flush before the MPI_Win_unlock but the program's
> behavior was not changed.
> The segmentation fault occured at the same MPI_Win_free.
>
> Best regards,
> Akihiro Tabuchi
>
>
> On 2015年11月05日 04:11, khaled hamidouche wrote:
>
>> Dear Akihiro,
>>
>> There seems to be some interaction between IPC and MPI_Put, we are
>> taking a look at this and get back to you.
>> But in meantime, can you please try to use Flush synchronization before
>> the unlock. This should help in fixing the behavior you are seeing.
>>
>> Thanks
>>
>>
>
> --
> Akihiro Tabuchi
> tabuchi at hpcs.cs.tsukuba.ac.jp
>
--
K.H
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20151105/3592d9e9/attachment-0001.html>
More information about the mvapich-discuss
mailing list