pnetcdf and mvapich2 2.2

[Wei-keng Liao]

Forgot to attach the program file.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: test_open_no_such_file.c
Type: application/octet-stream
Size: 1231 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/parallel-netcdf/attachments/20170203/6889a6c3/attachment.obj>
-------------- next part --------------


Wei-keng

On Feb 3, 2017, at 5:21 PM, Wei-keng Liao wrote:

> Hi, Mark
> 
> The Lustre driver in mvapich2 appears to append O_CREAT to the open mode,
> (line 50, in file src/mpi/romio/adio/ad_lustre/ad_lustre_open.c), even
> if the file is open for read-only. This is the root cause of one of the
> error messages you are seeing:
>   "expect error code NC_ENOENT but got NC_ENOTNC"
> 
> Attached is a small MPI program to verify such error. Could you please
> give it a try on your mvapich2 build and Lustre?
> Compile command:
>    mpicc test_open_no_such_file.c -o test_open_no_such_file
> Run command:
>    mpiexec -n 1 ./test_open_no_such_file /lustre/path/non-exist-file
> 
> If it is indeed an internal issue of mvapich, I can file a bug report to them.
> thanks
> 
> 
> Wei-keng
> 
> On Feb 3, 2017, at 12:25 PM, Wei-keng Liao wrote:
> 
>> Hi, Mark
>> 
>> For running "make check" on Lustre, could you please set the environment
>> variable PNETCDF_HINTS to "nc_header_align_size=512;nc_var_align_size=1"
>> and run "make check" again? I think it should pass make check. Do let me
>> know. These errors only occur for file systems whose striping size is
>> larger than 1. So, ext4 is not affected. I am working on a fix for that
>> test program. Please note this is a bug in the test program. the PnetCDF
>> library itself is intact.
>> 
>> When running "make check", I suggest not to set the environment variable
>> PNETCDF_VERBOSE_DEBUG_MODE, as many error checks are designed on
>> purpose. Those debugging messages can easily mask the true errors. That
>> environment variable is designed for testing one program at a time.
>> 
>> As for the errors from mvapich2, I do not have access to a machine with
>> infiniband and thus could not give it a try. However, the errors look like
>> a similar issue that has been discovered in OpenMPI recently: fail to
>> return the correct MPI error codes. I will look into the mvapich2 source
>> codes to confirm.
>> 
>> Thanks for trying various compilers and reporting the problem !
>> 
>> Wei-keng
>> 
>