Improper use of cache in parallel builds opening sequential files

[edwardhartnett]

While working on a fortran issue (Unidata/netcdf-fortran#253) I noticed this code in hdf5open.c:

#else /* only set cache for non-parallel. */
    if (H5Pset_cache(fapl_id, 0, nc4_chunk_cache_nelems, nc4_chunk_cache_size,
                     nc4_chunk_cache_preemption) < 0)
        BAIL(NC_EHDFERR);
    LOG((4, "%s: set HDF raw chunk cache to size %d nelems %d preemption %f",
         __func__, nc4_chunk_cache_size, nc4_chunk_cache_nelems,
         nc4_chunk_cache_preemption));
#endif /* USE_PARALLEL4 */

This is only setting the cache for non-parallel I/O builds.

It's true that the cache should not be used with parallel I/O, but parallel I/O builds can also do sequential opens. So cache should be applied as long as the file was not opened parallel, even if parallel I/O was built in. This should not be an ifdef, but an if statement.

I will address this after I finish with the fortran issue.

[edwardhartnett]

OK, I have a fix for this issue.

The issue was preventing HPC systems from benefiting from higher cache settings (see #1541).

Since the netCDF cache settings were not being used for parallel builds, HPC users will not see the benefits of the increased cache sizes, which is a bummer.

The PR I'm about to put up fixes that.