[Pvfs2-users] HDF5 & PVFS2

[Robert Latham]

On Fri, May 12, 2006 at 05:43:25PM +0200, Dries Kimpe wrote:
> I've been trying to get the HDF5 tests working, and they always fail
> when using pvfs2://.. files. (As opposed to not specifying any prefix at all)
> This data corruption does not occur when specifying ufs:// or nfs://
> (and actually running on nfs or local)

Hi Dries
I really haven't forgotten about you.  The HDF5 tests uncovered a few
bugs in ROMIO.  There is one last test failing that I need to fix, but
in the meantime, here's what works:

- Several HDF5 tests would create odd struct types that would cause
  MPICH-2 to die with a floating point exception.  MPICH2-1.0.4-rc1
  has the fix for this. 

- One test (cmpdsetr) tries to use POSIX routines on ROMIO-style file
  names.  I told the HDF5 guys about this, they identified it as a bug
  in their test, and fixed it on their end.  I presume the next
  release will have the fix.
 
- There was a bug in the PVFS2 listio processing.  I was not able to
  fix the code, but I did put logic in to detect the problematic
  condition and fall back to a much slower and correct method.  An
  imperfect version of this detection logic can also be found in
  MPICH2-1.0.4-rc1.  The next -rc will have an improved detection
  method.  At some point in the not-so-distant future we will be
  switching to a different approach for PVFS2 which will be less buggy
  and higher performance, but I wasn't confident in my ability to
  shake out all the bugs for the impending mpich2-1.4.0 release. 

- With all those issues fixed, there is still one last bug in ccchunkw
  with 3 processors.  I'm still working on understanding what's going
  on with that workload, and where the problem lies.  

Thanks for pointing out the issues in testphdf5.  It has proven to be
a very good test of ROMIO and PVFS2.  Sorry I didn't think to try it
out earlier.  

==rob

-- 

Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Labs, IL USA                B29D F333 664A 4280 315B