-
Notifications
You must be signed in to change notification settings - Fork 423
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
UCP/TAG: Move TM stuff from context to worker #2031
Conversation
Tag Matching queues and other related info are moved to the worker object. Now tag communications should not be crossed between different workers (like communicators in MPI). And ucp_tag_msg_recv_nb* behavior matches their description, that these routines receive a message on the particular worker.
Build finished. |
Test PASSed. |
Test PASSed. |
This is major change that we have discussed during F2F. I would suggest to get in touch with MPICH and OpenMPI community to make sure that - (a) they are happy with the changes , (b) we are not screwing anything |
It should be ok with Open MPI, where only one worker is used by UCX pml |
No need to get mutex in tag AM callbacks, since they can be called from progress context only and ucp_worker_progress is already guarded with locks.
Build finished. |
Test PASSed. |
Test FAILed. |
Test FAILed. |
the fault is #2027 |
bot:mlx:retest |
Test FAILed. |
Test FAILed. |
Conflicts: src/ucp/core/ucp_context.c src/ucp/core/ucp_context.h src/ucp/tag/offload.c
Build finished. |
Test PASSed. |
Test PASSed. |
@yosefe, plz check 2-nd commit. I removed unnecessary locks from tag AM handlers |
Tag Matching queues and other related info are moved to the worker object. Now tag communications should not be crossed between different workers (similar to communicators in MPI). And
ucp_tag_msg_recv_nb*
behavior matches their description, that these routines receive a message on the particular worker.