pnetcdf behavior?

[Shawn Kim]

Hi,

With your help, now I can trace IO related calls, MPI_File_write_all, in
Flash IO using pnetcdf. Flash IO (or pnetcdf) issues 50 MPI_File_write_all()
function calls, and I logged these calls with timestamps in MPI-level, PVFS2
client, and PVFS2 server. So I know which calls go where.

One weird thing is that there are missing calls in PVFS2 server logs. I got
50 IO calls in MPI-level and PVFS2 client log file, but only 38 calls in
PVFS server logs. I believe that my scheme can even trace collective IO. I
also know that Flash IO does not two phase collective buffering  and
noncontiguous access.
(The same things happened when using Flash IO with hdf5. Hdf5 issues 32
MPI_FIle_write_at_all() and some calls disappear.)

I post this question in pnetcdf, not PVFS2 forum because I think this is
caused by Flash IO or pnetcdf-1.1.1. Actually I tested my mpich2 and pvfs2
with a simple MPI program which uses the same MPI_File_write_all and got all
the correct results.

Do I miss something?

Regards,
Shawn
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/parallel-netcdf/attachments/20100314/1ac6c587/attachment.htm>