SPML/UCX: fixed hang in SHMEM_FINALIZE #6918
Conversation
hoopoepg
commented
Aug 21, 2019
hoopoepg
commented
- used MPI _Barrier to synchronize processes
- used MPI _Barrier to synchronize processes Signed-off-by: Sergey Oblomov <[email protected]>
hoopoepg
force-pushed
the
topic/fixed-hand-on-shmem-finalize
branch
from
2fde838 to
182023f
Compare
oshmem/mca/spml/ucx/spml_ucx.c
Outdated
| ret = opal_common_ucx_del_procs_nofence(del_procs, nprocs, oshmem_my_proc_id(), | ||
| mca_spml_ucx.num_disconnect, | ||
| mca_spml_ucx_ctx_default.ucp_worker); | ||
| /* Do not barrier here - barrier is called in _shmem_finalize */ |
There was a problem hiding this comment.
No need to barrier here - barrier is called in _shmem_finalize
| @@ -323,6 +321,8 @@ int mca_spml_ucx_add_procs(ompi_proc_t** procs, size_t nprocs) | |||
| free(wk_roffs); | |||
|
|
|||
| SPML_UCX_VERBOSE(50, "*** ADDED PROCS ***"); | |||
|
|
|||
| opal_common_ucx_mca_proc_added(); | |||
There was a problem hiding this comment.
there is detection of missing event handler. original call was from incorrect function
Signed-off-by: Sergey Oblomov <[email protected]>
hoopoepg
force-pushed
the
topic/fixed-hand-on-shmem-finalize
branch
from
42b33ad to
01dacaa
Compare
|
bot:retest |
|
bot:retest |
|
@yosefe ok to merge? |
|
@hoopoepg pls port to 4.0.x |
|
This PR should not have been merged as it was -- the commit message message on the first commit directly contradicts the commit content. Mistakes happen, but this particular pattern "make a commit, and then make another commit to fix the first commit" is a bit of a pet peeve of mine. Please make the final merged commits be good, high-quality commits whenever possible. The two commits on this PR should have been squashed before merging so that we didn't have to have a bad commit on master in the first place (which then gets propagated out to the release branches). Thanks. |
|
Oh, sorry I should have caught that in review. |
