[GH-ISSUE #323] Questions about double buffer in LL #225

New issue

Closed

opened 2026-10-26 09:56:56 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:56:56 +00:00

Owner

Originally created by @GitAlice123 on GitHub (Jul 22, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/323

I am curious about this picture. Say dispatch 0 using buffer 0, and dispatch 1 using buffer 1. I notice that when dispatch 1 (using buffer 1) issue, it will clean 'next buffer', which is buffer 0's signal buffer. Will it disturb dispatch 0's receive, which is polling the signal buffer 0?
And another question, combine 0's issue will use buffer 0. Can we make sure that when combine 0 issues, dispatch 0's issue has been completed? If we can't, combine 0 issue will rewrite the send buffer 0, which dispatch 0 might be still using it for sending.

Originally created by @GitAlice123 on GitHub (Jul 22, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/323 <img width="885" height="437" alt="Image" src="https://github.com/user-attachments/assets/fafc36f1-3d08-46e0-bb35-d200f08e6652" /> * I am curious about this picture. Say **dispatch 0** using **buffer 0**, and **dispatch 1** using **buffer 1**. I notice that when dispatch 1 (using buffer 1) issue, it will clean '**next buffer**', which is buffer 0's signal buffer. Will it disturb dispatch 0's receive, which is polling the signal buffer 0? * And another question, combine 0's issue will use buffer 0. Can we make sure that when combine 0 issues, dispatch 0's issue has been completed? If we can't, combine 0 issue will rewrite the send buffer 0, which dispatch 0 might be still using it for sending.

bsakaguchi closed this issue

2026-10-26 09:56:56 +00:00

bsakaguchi commented

2026-10-26 09:56:58 +00:00

Author

Owner

@sphish commented on GitHub (Aug 8, 2025):

Sorry for the late reply. The picture in https://github.com/deepseek-ai/profile-data/blob/main/README.md may be more clear

In fact, there won’t be two communications running simultaneously on the same GPU.

@sphish commented on GitHub (Aug 8, 2025): Sorry for the late reply. The picture in https://github.com/deepseek-ai/profile-data/blob/main/README.md may be more clear <img width="2584" height="618" alt="Image" src="https://github.com/user-attachments/assets/b86c11c5-ab19-4908-8239-5a5f9248a63c" /> In fact, there won’t be two communications running simultaneously on the same GPU.

bsakaguchi referenced this issue

2026-10-26 10:01:13 +00:00

[PR #225] [CLOSED] Refactor and support TMA for internode kernels #476