[mpich-discuss] Flash IO Benchmark using mpich2 and pvfs2

[Rob Latham]

On Wed, Mar 10, 2010 at 02:47:09PM -0500, Shawn Kim wrote:
> 
> 1. MPI_File_write_all (mpich2-1.0.8/src/mpi/romio/mpi-io/write_all.c)
> 2. MPIOI_File_write_all (mpich2-1.0.8/src/mpi/romio/mpi-io/write_all.c)
> 3. ADIO_WriteStridedColl
> 4. ADIOI_GEN_WriteStridedColl
> (mpich2-1.0.8/src/mpi/romio/adio/common/ad_write_coll.c)
> -------------------------------------------------------------------------------------------------------------------------------------
> >From here, I can't trace it
> 5. ADIO_WriteContig (ad_pvfs2_write.c)
> 6. ADIOI_PVFS2_WriteContig
> 7. PVFS_sys_write
> ====================================
> 8. PVFS_sys_io (sys-io.sm)
> 
> Is there anybody who knows why these call sequences don't work? If anybody
> ran Flash IO before, is the result correct?
> Or, am I missing something?

Hi: this looks correct to me.  what are you confused about?

if the I/O accesses overlap, or the "romio_cb_write" hint is set to
"enable", "ADIOI_GEN_WriteStridedColl will carry out the two-phase
collective I/O optimization. 

If you are jumping to ADIO_WriteContig, then you might have hit this
test fairly early in ADIOI_GEN_WriteStridedColl:

   if (fd->hints->cb_write == ADIOI_HINT_DISABLE ||
           (!interleave_count && (fd->hints->cb_write == ADIOI_HINT_AUTO))) {

if that evaluates true, then you fall into "process this request
independently".

Otherwise, ADIOI_Exch_and_write and ADIOI_W_Exchange_data contain the
heart of the two-phase optimization.  Phase one is exchanging data with
processors to re-order the I/O, and Phase two is the actual
(contiguous) write, or ADIO_WriteContig.

==rob

> Thank you
> 
> - Best regards,
> Shawn

> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss


-- 
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA