MPI synchronous send does not guarantee data placement into the target
buffer. It only guarantees that the recv operation on the target side
has been posted and matched. The data can still come later.
Ken
On 08/02/2018 05:20 PM, Hefty, Sean wrote:
Based on discussion with some MPI developers, I'm proposing that we introduce a
new completion semantic to libfabric. This specifically targets support for
the MPI synchronous send operation.
The proposed semantic is:
FI_MATCH_COMPLETE*
: Indicates that a completion should be generated only after the operation
has been placed into the application specified buffer. Operations using
this completion semantic are dependent on the application at the target
claiming the message or results. As a result, match complete may involve
additional provider level acknowledgements or lengthy delays. However, this
completion model enables peer applications to synchronize their execution.
The associated PR is:
https://github.com/ofiwg/libfabric/pull/4303
Note that the PR consolidates and expands the completion levels supported by
libfabric into the fi_cq man page.
- Sean
_______________________________________________
ofiwg mailing list
[email protected]
https://lists.openfabrics.org/mailman/listinfo/ofiwg
_______________________________________________
ofiwg mailing list
[email protected]
https://lists.openfabrics.org/mailman/listinfo/ofiwg
