[mvapich-discuss] open non-existing files on Lustre expecting MPI_ERR_NO_SUCH_FILE but got MPI_SUCCESS

[Wei-keng Liao]

Hi, Hari

I do not have an access to a machine with infiniband and Lustre.
I have forwarded your patch to Mark Dixon and he will give it a try
and get back to you soon. Please keep him in the loop. Thanks.


Wei-keng

On Feb 7, 2017, at 4:09 PM, Hari Subramoni wrote:

> Hello Wei-keng,
> 
> Can you please try the attached patch and see if things work for you?
> 
> This patch will be available by default with future releases of the MVAPICH2 library.
> 
> Best Regards,
> Hari.
> 
> On Tue, Feb 7, 2017 at 4:50 AM, Hari Subramoni <subramoni.1 at osu.edu> wrote:
> Hello Wei-keng, 
> 
> Many thanks for the report. We will take a look at it and get back to you soon. 
> 
> Best Regards, 
> Hari.
> 
> On Feb 7, 2017 1:12 AM, "Wei-keng Liao" <wkliao at eecs.northwestern.edu> wrote:
> Hi,
> 
> I am a developer of PnetCDF library and would like to file a bug report.
> Mark Dixon, a PnetCDF user cc-ed in this email, reported an error when
> building PnetCDF with mvapich2-2.2. The root cause of the error is due to
> mvapich2-2.2 fails to return the expected MPI error class MPI_ERR_NO_SUCH_FILE
> when a test program tries to open a non-existing file on Lustre.
> 
> After some diggings, I found out that the Lustre driver in mvapich2 appends
> O_CREAT flag to the mode argument of the open call, (line 50, in file
> src/mpi/romio/adio/ad_lustre/ad_lustre_open.c). Because of that, the file is
> mistakenly created and MPI_SUCCESS is returned, which is an unexpected outcome
> by MPI standard.
> 
> Attached is a small MPI program to reproduce such error. Mark has used it to
> verify the issue described above. He also ran it against an ext4 (UFS) and the
> correct error class was returned. So the problem is specific for Lustre driver.
> Please see the discussion email threads:
> http://lists.mcs.anl.gov/pipermail/parallel-netcdf/2017-February/001897.html
> 
> 
> Wei-keng
> 
> 
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
> 
> 
> <lustre_open.patch>