Skip to content

Enable DMA-BUF-backed RDMA memory registration - #3

Draft
JimyMa wants to merge 1 commit into
mainfrom
feat/dmabuf-rdma
Draft

Enable DMA-BUF-backed RDMA memory registration#3
JimyMa wants to merge 1 commit into
mainfrom
feat/dmabuf-rdma

Conversation

@JimyMa

@JimyMa JimyMa commented Aug 20, 2026

Copy link
Copy Markdown
Owner

Summary

  • add DMA-BUF-backed memory-region registration to the RDMA memory pool and Python bindings
  • preserve explicit IOVA values for local SGEs and published remote MR metadata
  • allow PeerAgent to own, lazily materialize, and clean up DMA-BUF regions without importing CUDA
  • add cross-process registration and three-process GPU-to-GPU RDMA write examples

Validation

  • pytest -q tests/python/test_directed_connection_io.py tests/python/test_rdma_memory_pool_unregister.py (19 passed)
  • python examples/python/p2p_rdma_dmabuf_cross_process.py --rdma-device mlx5_0 --link-type IB
  • python examples/python/p2p_rdma_dmabuf_three_process.py --source-cuda-device 0 --destination-cuda-device 1 --rdma-device mlx5_0 --link-type IB (16 MiB verified)

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.

1 participant