[Darshan-users] no write information when using mvapich2
Phil Carns
carns at mcs.anl.gov
Thu Jul 25 22:19:20 CDT 2013
Just a quick follow up for mailing list archival purposes. We discussed
this problem a little bit more off-list and it looks like the issue is
related to tracking AIO operations issued by the nonblocking I/O
functions in MPI-IO. Here is a trac ticket with some more information:
http://trac.mcs.anl.gov/projects/darshan/ticket/115
I don't know why this is being triggered in mvapich but not openmpi in
David's case, but I'm able to reproduce it with a recent mpich build as
well. We'll continue to look into it further.
thanks,
-Phil
On 07/19/2013 01:06 PM, David Shrader wrote:
> Hello All,
>
> I am currently trying to bring up a Darshan build that supports
> mvapich2 1.7 and 1.8. The build seemingly completes without an issue,
> but I have noticed a problem when running an IO-replication
> application that we like to use: I am getting no stats in terms of
> writes. The app's general flow is to do a writing phase and then a
> reading phase to a single file (no collective operations except when
> opening the file). The stats for the reading phase are about what I
> expect. However, the stats for the writes are all 0. Specifically,
> CP_BYTES_WRITTEN and CP_POSIX_WRITES are both 0. The corresponding
> READS counters are where they should be.
>
> I have tested this app (with the same run command) using a build of
> Darshan that works with Open MPI. There, the WRITE-related counters I
> mentioned above are where I expect them to be. It is just the mvapich2
> builds that have this behavior.
>
> Did I miss something when building for mvapich2? Any ideas as to why
> write information might be missing? If it helps, I can send some of
> the output from darshan-parser (there is only a single file being
> written to, so the output isn't that large) or the log file itself.
>
> Thanks,
> David
>
More information about the Darshan-users
mailing list