OSC RDMA initialization cleanups #9835
Merged
Conversation
This file contains hidden or 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
wzamazon
reviewed
Jan 5, 2022
wzamazon
reviewed
Jan 5, 2022
wzamazon
reviewed
Jan 5, 2022
Add ofi to the list of btls that do not need to verify connectivity, since the most common use case of OFI does provide this property. Rename the list of btls with full connectivity from ompi_osc_rdma_btls to ompi_osc_rdma_full_connectivity_btls to make it a little more clear what the list represents. Signed-off-by: Brian Barrett <[email protected]>
When the rdma and pt2pt onesided components both existed, there was a use case for lowering the priority of the RDMA component when certain MTLs are in use. With the pt2pt component removed, this is dead code. Signed-off-by: Brian Barrett <[email protected]>
With the removal of the pt2pt component, there's really not an advantage to having 2 levels of priority (depending on whether we have a true rdma btl or emulated), so go with a single priority. Signed-off-by: Brian Barrett <[email protected]>
The RDMA OSC component requires remote completion for its synchronization algorithms to work properly. This is phase one of two to handle BTLs which do not provide remote completion. This patch requires "accelerated" BTLs to support remote completion. Accelerated BTLs are those which can communicate with all peers in the Window (so multiple BTLs per Window are not used), natively provide the required put/get/atomic operations, and which can register suitable memory for the window's communication. Signed-off-by: Wei Zhang <[email protected]> Signed-off-by: Brian Barrett <[email protected]>
No functional changes (an initialization was reordered to make documentation easier), but add documentation about BTL selection behavior. Signed-off-by: Brian Barrett <[email protected]>
Skipping the over-optimization of supporting a BTL that can talk native RDMA to everyone but self (we don't have one of those today) allows a large simplification of the accelerated BTL selection on the OSC RDMA component. Signed-off-by: Brian Barrett <[email protected]>
bwbarrett
force-pushed
the
bugfix/rdma-osc-accelerated-path
branch
from
2630f84 to
e0a5b8d
Compare
|
Updated with the commit message cleaned up and the comment cut-n-paste issue removed. |
wzamazon
approved these changes
Jan 5, 2022
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 is the first chunk of PRs aimed at fixing #8983. This will ensure that "accelerated" BTLs (those that have native RDMA and can reach all peers) also provide the required remote completion. This is not a complete fix for #8983, as the alternate BTL path in the osc rdma component still has a remote completion failure mode. Patches for that will be coming as a follow-on patch series.