[mpich-discuss] extra bytes

Kokron, Daniel S. (GSFC-610.1)[Computer Sciences Corporation] daniel.s.kokron at nasa.gov
Thu Aug 9 14:33:16 CDT 2012

The attached program writes a single distributed 3D field to a file using calls to MPI_Type_create_subarray,
MPI_File_set_view and MPI_File_write_all.  It then attempts to append two 1D variables to the end of the file
from a single rank.  I've implemented three scenarios for the 1D writes.  See the case structure for scenarios.

The final size of the file should be [(516x516x72)+(2*72)]*4 = 76682304 Bytes.
This is the case under scenario 0 and 1 for all process counts (isize*jsize) I've tested.
I get the correct size under scenario 2 for all process counts tested only
if isize is 3 or smaller.
With isize=4, I get a file that is 76683852 B (+1548B)
isize=5 -> 76683952  (+100B)
isize=6 -> 76684024   (+72B)
isize=7 -> 76684072   (+48B)
isize=8 -> 76685912 (+1840B)
isize=9 -> 76685968   (+56B)

These results are from using the Intel compiler with SGI MPT-2.06 on the pleiades platform (Linux
I've also confirmed that the odd behavior occurs under MPT-1.25, IntelMPI-, IntelMPI-3.1.038,
MVAPICH2-1.8 and OpenMPI-1.5.5 on the same platform with the same compiler.

The odd behavior is present when writing to Lustre or a RAM FS on pleiades.

The odd behavior is present on the discover platform (Linux using Intel with IntelMPI- and writing to a GPFS file system.

Is this expected behavior?

Daniel Kokron
SciCon group
-------------- next part --------------
A non-text attachment was scrubbed...
Name: tst_2Dlayout.F90
Type: application/octet-stream
Size: 14047 bytes
Desc: tst_2Dlayout.F90
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20120809/e8cc5fdd/attachment.obj>

More information about the mpich-discuss mailing list