Topo aware coll comm v2 #12032
Merged
Topo aware coll comm v2 #12032
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
bosilca
reviewed
Oct 30, 2023
ompi/communicator/comm_cid.c
Outdated
| static int ompi_comm_activate_complete (ompi_communicator_t **newcomm, ompi_communicator_t *comm) | ||
| /* Callback function to set comunicator disjointness flags */ | ||
| static inline void ompi_comm_set_disjointness_nb_complete (ompi_comm_cid_context_t *context) { | ||
| if (!OMPI_COMM_IS_DISJOINT_SET(*context->newcommp)) { |
There was a problem hiding this comment.
This function should only be called once per communicator. Calling a second time should trigger an error, at least a message in debug mode.
There was a problem hiding this comment.
Thanks I will think about that.
There was a problem hiding this comment.
Please see update - I decided to add a warning message instead of hard failure.
ompi/communicator/comm_cid.c
Outdated
| * send messages over the new communicator | ||
| /** | ||
| * Dual-purpose barrier: | ||
| * 1. The communicator's disjointness is inferred fromimax_local_peers. |
wenduwan
force-pushed
the
topo_aware_coll_comm_v2
branch
from
c24c3e8
to
ea62e7c
Compare
|
@bosilca Heartbeat - I imagine you are busy in SC23. Please kindly review the change when you have time. |
bosilca
approved these changes
Dec 2, 2023
bosilca
reviewed
Dec 2, 2023
This patch introduces a new communicator flag to indicate if no processes share the same node. This flag is intended for optimization of hierarchy-aware collective operations to select the more efficient transport/algorithm. In this change we introduce a non-blocking allreducestep in communicator activation and set the new flag in the completion callback function. Signed-off-by: Wenduo Wang <[email protected]>
This patch introduces assertions to verify that sub-communicators are created with the expected OMPI_COMM_DISJOINT* flags. Signed-off-by: Wenduo Wang <[email protected]>
wenduwan
force-pushed
the
topo_aware_coll_comm_v2
branch
from
4fcafb0
to
61c9403
Compare
|
@bosilca Thank you very much! I have fixed the typo and rebased the branch. |
bosilca
approved these changes
Dec 2, 2023
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch refactors #11864 and prevents the deadlock reported in #12016
Specifically, this patch breaks down the disjointness flag determination logic in 2 parts:
max_local_peersin a (non-blocking) allreduce unconditionally, and thenmax_local_peersvalue and set the disjointness flag.As a result, this change makes the
context->idfield obsolete, and therefore we simply remove it. The allreduce also serves as a barrier function so we always invoke it, so we do not need the previous logic to check the parent disjointness flag - this simplifies the logic.I ran the reproducer provided by @dalcinl
NOTE: I observed a deadlock behavior while making this change:
context->idandcontext->max_local_peers, and pass them assubreqstoompi_comm_request_schedule_append (request, ompi_comm_activate_nb_complete, &subreqs, 2);the reproducer will hang. Back trace shows that the allreduce never completes:The allreduce used in this case was
ompi_comm_allreduce_intra_pmix_nb