feat(libfabric): Implement notification fragmentation for large messages #1142
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
akkart-aws
requested review from
amitrad-aws,
fengjica and
yexiang-aws
as code owners
|
👋 Hi akkart-aws! Thank you for contributing to ai-dynamo/nixl. Your PR reviewers will review your contribution then trigger the CI to test your changes. 🚀 |
akkart-aws
changed the title
akkart-aws
force-pushed
the
increase_notif_size
branch
from
1fab141 to
a1c185a
Compare
akkart-aws
force-pushed
the
increase_notif_size
branch
2 times, most recently
from
b4ae63a to
dd24c6d
Compare
fengjica
reviewed
Dec 23, 2025
Collaborator
fengjica
left a comment
fengjica
left a comment
There was a problem hiding this comment.
Hi, I'm just sending out a partial review for now, and will keep reviewing it later.
fengjica
reviewed
Dec 23, 2025
amitrad-aws
requested changes
Dec 25, 2025
Collaborator
amitrad-aws
left a comment
amitrad-aws
left a comment
There was a problem hiding this comment.
I think some of the intermidiate commits here are misleading in that they change the behavior to something we wouldn't want to do (even as an intermidiate change).
I would squash "add notification fragmentation...", "add attribute((packed))...", "auto-calculate fragment size..." into one commit.
akkart-aws
force-pushed
the
increase_notif_size
branch
4 times, most recently
from
08dc93b to
87ba2c6
Compare
akkart-aws
force-pushed
the
increase_notif_size
branch
from
87ba2c6 to
8f7edaf
Compare
amitrad-aws
requested changes
Dec 30, 2025
akkart-aws
force-pushed
the
increase_notif_size
branch
from
8f7edaf to
629b15a
Compare
fengjica
previously approved these changes
Jan 2, 2026
akkart-aws
force-pushed
the
increase_notif_size
branch
2 times, most recently
from
f8f6e9c to
ccde7f9
Compare
- Update comments to accurately reflect request tracking initialization and notification handling - Change `pending_notifications_` from std::map to std::unordered_map for O(1) lookup performance - Improve inline documentation for PendingNotification struct fields These changes improve code readability and maintainability without altering functionality.
- Use Emplace for map operations to avoid additional copies. - Use passbyreference in fragmentNotificatioMessage to prevent vector copy. Signed-off-by: Arun Karthik <[email protected]>
…ders Some of the notification fragment header fields are common across all the notification fragments. This commit takes such fields and puts them into a spearate header for first notification fragment. This commit also moves the agent name from the header and combines it with a common notification buffer.
Signed-off-by: Arun Karthik <[email protected]>
akkart-aws
force-pushed
the
increase_notif_size
branch
from
2f9f3ea to
561c0a5
Compare
akkart-aws
force-pushed
the
increase_notif_size
branch
from
561c0a5 to
b2b6849
Compare
Contributor
|
/build |
Contributor
|
/ok to test b2b6849 |
fengjica
approved these changes
Jan 6, 2026
amitrad-aws
approved these changes
Jan 7, 2026
ovidiusm
pushed a commit
to ovidiusm/nixl
that referenced
this pull request
Jan 8, 2026
…ges (ai-dynamo#1142) * fix(libfabric): Improve hexDump * refactor(libfabric): Move xfer_id allocation from request pool to caller * refactor(libfabric): move xfer_id from BinaryNotification to backend handle * refactor(libfabric): simplify pending notification map logic and fix constructor * refactor(Libfabric): Modify sendNotifPriv to use vector of BinaryNotif * refac(libfaric): unify notif processing for all expected_completions values * feat(libfabric): add notification fragmentation on sender and reassembly on receiver * refactor(libfabric): rename method names and some code clarity - Update comments to accurately reflect request tracking initialization and notification handling - Change `pending_notifications_` from std::map to std::unordered_map for O(1) lookup performance - Improve inline documentation for PendingNotification struct fields These changes improve code readability and maintainability without altering functionality. * rev(libfabric): Fix comments on the notification refactor - Use Emplace for map operations to avoid additional copies. - Use passbyreference in fragmentNotificatioMessage to prevent vector copy. Signed-off-by: Arun Karthik <[email protected]> * rev(libfabric): Split notification fragments header into separate headers Some of the notification fragment header fields are common across all the notification fragments. This commit takes such fields and puts them into a spearate header for first notification fragment. This commit also moves the agent name from the header and combines it with a common notification buffer. * rev(Libfabric): Remove Submitted requests as function callback Signed-off-by: Arun Karthik <[email protected]> * fix(libfabric): Update copyright year --------- Signed-off-by: Arun Karthik <[email protected]>
nv-nmailhot
pushed a commit
that referenced
this pull request
Jan 12, 2026
…ges (#1142) (#1182) * fix(libfabric): Improve hexDump * refactor(libfabric): Move xfer_id allocation from request pool to caller * refactor(libfabric): move xfer_id from BinaryNotification to backend handle * refactor(libfabric): simplify pending notification map logic and fix constructor * refactor(Libfabric): Modify sendNotifPriv to use vector of BinaryNotif * refac(libfaric): unify notif processing for all expected_completions values * feat(libfabric): add notification fragmentation on sender and reassembly on receiver * refactor(libfabric): rename method names and some code clarity - Update comments to accurately reflect request tracking initialization and notification handling - Change `pending_notifications_` from std::map to std::unordered_map for O(1) lookup performance - Improve inline documentation for PendingNotification struct fields These changes improve code readability and maintainability without altering functionality. * rev(libfabric): Fix comments on the notification refactor - Use Emplace for map operations to avoid additional copies. - Use passbyreference in fragmentNotificatioMessage to prevent vector copy. * rev(libfabric): Split notification fragments header into separate headers Some of the notification fragment header fields are common across all the notification fragments. This commit takes such fields and puts them into a spearate header for first notification fragment. This commit also moves the agent name from the header and combines it with a common notification buffer. * rev(Libfabric): Remove Submitted requests as function callback * fix(libfabric): Update copyright year --------- Signed-off-by: Arun Karthik <[email protected]> Co-authored-by: Arun Karthik <[email protected]> Co-authored-by: Mikhail Brinskiy <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
What?
Implemented notification message fragmentation in the NIXL libfabric backend to support unlimited notification sizes by splitting messages into 256-byte chunks.
Why?
TensorRT-LLM's disaggregated serving architecture requires exchanging connection metadata between nodes. On P5 instances with 32 EFA devices, this connection data is approximately 1,792 bytes (32 devices × 56 bytes per endpoint), which exceeds the previous 1,024-byte notification limit, causing connection establishment to fail.
This change enables TensorRT-LLM to work on P5 instances and scales to support 256+ rails per node (~14KB connection data).
How?
Sender-Side:
prepXfer(), splitting messages into 256-byte chunksnotifSendPriv()sends all fragments with sequence metadataReceiver-Side:
checkPendingNotifications()reassembles complete messagesAdditional Improvements:
__attribute__((packed))to BinaryNotification for cross-platform compatibilityfragmentNotificationMessage()helper function for code reusabilityModified Files:
src/utils/libfabric/libfabric_common.hsrc/plugins/libfabric/libfabric_backend.hsrc/plugins/libfabric/libfabric_backend.cppsrc/utils/libfabric/libfabric_rail.cpp