bug fix: PnetCDF vard APIs when filetype is MPI_DATATYPE_NULL

[wkliao]

When argument filetype is MPI_DATATYPE_NULL in vard APIs, it indicates this request is of zero length. Thus we can skip updating the counters.

Failed test program:
https://github.com/Parallel-NetCDF/PnetCDF/blob/master/test/testcases/test_vard_multiple.c

[shanedsnyder]

We've actually run into similar issues in the past with the MPI-IO module, not sure why it didn't occur to me to error check this.

Instead of bailing out of instrumentation, I think we should instrument as normal, just marking the access size as 0. That matches what we do in MPI-IO, and would allow us to detect that applications are doing 0-byte operations. Could we instead add logic to CALC_VARD_ACCESS_INFO macro to just set access size to 0 in cases where the given datatpye is NULL?

[wkliao]

revised as suggested

[shanedsnyder]

Thanks, looks good from my local testing.

Quick question I noticed when sanity checking the output log for the test case that was failing: What would we expect the total number of bytes to be transferred from this portion of the test to be:

    /* bufcount is ignored when buftype is MPI_DATATYPE_NULL */
    err = ncmpi_put_vard_all(ncid, varid[0], MPI_INT, buf[0], -1,
                             MPI_DATATYPE_NULL); CHECK_ERR

0 bytes, as well, since the memory datatype is MPI_DATATYPE_NULL?

[wkliao]

In this case, the write amount is 4 bytes, i.e. size of 3rd argument, filetype.
The buftype being MPI_DATATYPE_NULL means the write buffer is contiguous and its amount is equal to the size of filetype.

[shanedsnyder]

Perfect, that matches what Darshan is instrumenting so all looks good.

Thanks, I'll go ahead and merge this.