Skip to content
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

RDMA Executor support #147

Closed
wants to merge 70 commits into from

Conversation

mustafabar
Copy link
Contributor

Added

  1. Support for RDMA-based data transfers in TransferBench. This executor enables data transfers using RDMA-capable NICs, bypassing host involvement for GPU/CPU transfers
  2. Two new executor symbols are added: I and N. Nj is the nearest NIC to GPUj. Ij is NIC with index j.
  3. Memory is preregistered so that the cost of registration is not included in the transfer
  4. Added topology display support using PCIe tree mapping algorithm
  5. Multi-port NIC mapping support so that GPUs that share a PCIe bridge with a GPU are evenly distributed on the ports
  6. QPs are expressed as the number of sub-executors per transfer. Data is split over the QPs
  7. Automatic GID Index detection for RoCE traffic (Code extracted from recent NCCL support)
  8. When topology detection is not accurate, especially in systems with front NICs and multiport NICs, the CLOSEST_NIC variable can override the automatically detected topology

@mustafabar mustafabar requested a review from a team as a code owner December 2, 2024 16:40
@mustafabar mustafabar requested review from a team and removed request for a team December 2, 2024 16:40
@mustafabar mustafabar self-assigned this Dec 2, 2024
mustafabar and others added 27 commits December 4, 2024 14:26
* Adding NIC executor support

* Removing debug compile flag

* Updating CHANGELOG, fixing subiterations, updating topology display, adding NICs to a2a preset

* Adjusting minimum bandwidth calculation for a2a to account for NIC

* Adding relaxed ordering for NIC

* Minor variable name fix to separate memreg vs init flags

---------

Co-authored-by: Gilbert Lee <[email protected]>
@mustafabar mustafabar requested a review from a team as a code owner January 14, 2025 15:33
@gilbertlee-amd
Copy link
Collaborator

Duplicate of v1.59 candidate

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants