IBGDA companion QP + counter infrastructure#989
Closed
dmwu wants to merge 3 commits intometa-pytorch:mainfrom
Closed
IBGDA companion QP + counter infrastructure#989dmwu wants to merge 3 commits intometa-pytorch:mainfrom
dmwu wants to merge 3 commits intometa-pytorch:mainfrom
Conversation
meta-cla
bot
added
the
CLA Signed
Summary: Pull Request resolved: meta-pytorch#987 Differential Revision: D94633071
Summary: [pipes-ibgda] Signal ownership: move signal buffers from transport to caller Refactor the IBGDA transport layer so that signal buffers are owned by the window layer (HostWindow/DeviceWindow), not by P2pIbgdaTransportDevice. This separation is foundational for the unified dual NVL+IBGDA window design, where the window allocates, RDMA-registers, and exchanges all signal/counter buffers across peers, while the transport provides only QPs for RDMA operations. Changes to P2pIbgdaTransportDevice: - Remove signal buffer members; all signal APIs (put_signal, wait_signal, read_signal, reset_signal) now take explicit IbgdaLocalBuffer/IbgdaRemoteBuffer arguments so the caller controls buffer lifetime and placement. - reset_signal() rewritten as an RDMA inline write (doca_gpu_dev_verbs_p<uint64_t>) to a remote signal buffer, replacing the old local volatile store. Signals are only ever updated by remote RDMA atomics, so resetting must also go through RDMA. - Add kDefaultDeviceTimeoutCycles (~5-7s on typical GPU clocks) as a compile-time default for device-side wait operations to prevent indefinite hangs. Callers updated: - MultipeerIbgdaDeviceTransport, MultipeerIbgdaTransportCuda: pass explicit signal buffers through kernel launch wrappers. - MultipeerIbgdaTransportTest: all existing test cases updated to allocate and pass signal buffers explicitly. - IbgdaBenchmark: all existing benchmarks updated to pass signal buffers. Reviewed By: siyengar Differential Revision: D94937310
Summary: Add companion QP and counter infrastructure to the IBGDA transport layer. Each peer gets a companion QP (core_direct=true for WAIT WQE support) used for NIC completion tracking via loopback RDMA atomics. A zero-based sink buffer (ibv_reg_mr_iova2 with iova=0) discards RDMA atomic return values. Device-side additions to P2pIbgdaTransportDevice: - signal_remote(): RDMA atomic to caller-provided remote signal buffer - signal_remote_with_fence(): same with IBV_SEND_FENCE for data ordering - put_signal_counter_remote(): compound data + signal + counter operation - signal_counter_remote(): signal + counter without data write - wait_local(work, timeout): CQ poll at specific WQE index with optional timeout - All signal/counter buffer addresses are caller-provided (window-owned) Reviewed By: siyengar Differential Revision: D94937308
Contributor
|
This pull request has been merged in ecf383a. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Summary:
Add companion QP and counter infrastructure to the IBGDA transport layer. Each peer gets a companion QP (core_direct=true for WAIT WQE support) used for NIC completion tracking via loopback RDMA atomics. A zero-based sink buffer (ibv_reg_mr_iova2 with iova=0) discards RDMA atomic return values.
Device-side additions to P2pIbgdaTransportDevice:
Reviewed By: siyengar
Differential Revision: D94937308