nc_put_vars_double fails in parallel if using netcdf-4 collective

[gsjaardema]

Version 4.5.1-devel.

If I call nc_put_vars_double with stride=1 in parallel with some processors having no data to write, then the H5Dwrite call will fail.

The problem is due to the

   if(nels == 0)
      return NC_NOERR; /* cannot write anything */

at line 244 of libdispatch/dvarput.c. If I remove that early return and if stride == 1, then the code will complete correctly. If that line is left as is, then some processors return early and the code hangs down below H5Dwrite due to hdf5 calling PMPI_Allreduce if using collective io.

There is another issue if stride != 1, but I will report that in a separate issue.

[gsjaardema]

The related issue when stride != 1 is #448

[edhartnett]

@gsjaardema Is this still and issue or should it be closed?

[gsjaardema]

This can be closed.

[edhartnett]

You have to close it, I can't. ;-)