PML/UCX: fix problem with (i)mrecv #13620
Conversation
the communicator in the message struct was not being added to the returned request. Related to issue open-mpi#13484 Signed-off-by: Howard Pritchard <[email protected]>
|
@devreal see if this helps for ASAN. Note that i've not usually tried to run mpi4py with UCX PML. You have to skip test_ulfm among other things. |
|
Thanks @hppritcha! It seems to fix the issue. I wonder why we don't need to retain the communicator though. It could be released by the application once they got the message object. Or is the communicator retained already? I opened #13623 for the pml/ucx ULFM issue. We probably should start running the mpi4py test suite with UCX as well. |
|
UCX PML doesn't seem to like retaining/releasing the comm associated with a request. See for example the mca_pml_ucx_irecv function. a request object is "returned" from the ucp_tag_recv_nb UCX method. |
|
ignoring non-functional NVIDIA CI |
|
@hppritcha does this need to be cherry-picked to v5.0? |
|
yes I think it should. on my to-do list! |
4 participants
the communicator in the message struct was not being added to the returned request.
Related to issue #13484