James,
for the v1.8/v1.10 series, the fix is available at
https://github.com/ggouaillardet/ompi-release/commit/c301bab8c9aff76eb7a3ee56b965b6ff3cf0073c.diff
fwiw
- i ran the test program under the debugger, and the datatype is the
same before and after MPI_Type_create_resize (e.g. the compiler did not
add extra padding for alignment)
- if you need MPI_Type_create_resize to handle some alignment, it
*might* be more efficient to manually
add some padding so everything is well aligned *and* send the padding
with the rest of the message
(so you only send/recv a contiguous block of data). there is a
trade-off between the size of the data to be
sent, and the complexity of the datatype that requires more cpu and
less efficient internal memory copy.
Cheers,
Gilles
On 1/28/2016 10:11 AM, Gilles Gouaillardet wrote:
James,
that is definetly a bug in OpenMPI, and i will push a fix shortly
(well, at least when github is back online ...)
Thanks and regards,
Gilles
On 1/28/2016 1:46 AM, Ramsey, James J CIV USARMY RDECOM ARL (US) wrote:
In case the last message came out garbled because newlines were removed, I've send it
again. Anyway,I've been trying to create a derived datatype for use in one-sided
communication, and I find that attempting to modify the extents to account for potential
alignment issues is making things *worse*, at least when one-sided communication is used.
Modifying extents seems to work fine when using the datatype for two-sided communication,
however. I also find that the problem seems to occur with the version of OpenMPI on my
workstation (which, unfortunately, is a prior stable release, 1.8.5), but not with Cray's
MPICH. Mainly, I'm trying to figure out if there's a problem with my code that OpenMPI is
catching but Cray is not, or if the problem is at OpenMPI's end. The attached test code
that triggers the problem for me when compiled with "-DADJ_EXTENT".
_______________________________________________
users mailing list
us...@open-mpi.org
Subscription:http://www.open-mpi.org/mailman/listinfo.cgi/users
Link to this
post:http://www.open-mpi.org/community/lists/users/2016/01/28392.php
_______________________________________________
users mailing list
us...@open-mpi.org
Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
Link to this post:
http://www.open-mpi.org/community/lists/users/2016/01/28394.php
