deepseek-ai/DeepEP

Fork 0

mirror of https://github.com/deepseek-ai/DeepEP.git synced 2026-11-05 00:47:32 +00:00

[GH-ISSUE #147] Question about nvshmem.patch #129

New issue

Closed

opened 2026-10-26 09:54:55 +00:00 by bsakaguchi · 4 comments

bsakaguchi commented

2026-10-26 09:54:55 +00:00

Owner

Originally created by @LUIGINOVA5 on GitHub (May 7, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/147

Hi,
I noticed that after commit "Low latency kernels use rdma atomic to support AR.", ll no longer polls rx_wq during receive operations. Given this change, is regular recv queue from patch still necessary?
Besides, could you explain "changing QP creating order" improves ll performance? My understanding is that, it reduces the rc establishment time by distributing the initialization load across PEs, thereby preventing concurrent bottlenecks. Is this accurate?
Thank you!

Originally created by @LUIGINOVA5 on GitHub (May 7, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/147 Hi, I noticed that after commit "Low latency kernels use rdma atomic to support AR.", ll no longer polls rx_wq during receive operations. Given this change, is regular recv queue from patch still necessary? Besides, could you explain "changing QP creating order" improves ll performance? My understanding is that, it reduces the rc establishment time by distributing the initialization load across PEs, thereby preventing concurrent bottlenecks. Is this accurate? Thank you!

bsakaguchi closed this issue

2026-10-26 09:54:55 +00:00

bsakaguchi commented

2026-10-26 09:54:56 +00:00

Author

Owner

@sphish commented on GitHub (May 9, 2025):

Yes, recv queue can be removed.
Changing the QP creation order is intended to make the NIC's send behavior more balanced.

@sphish commented on GitHub (May 9, 2025): 1. Yes, recv queue can be removed. 2. Changing the QP creation order is intended to make the NIC's send behavior more balanced.

bsakaguchi commented

2026-10-26 09:54:56 +00:00

Author

Owner

@LUIGINOVA5 commented on GitHub (May 9, 2025):

Yes, recv queue can be removed.

Changing the QP creation order is intended to make the NIC's send behavior more balanced.

Thanks for reply! Regarding answer 2, earlier-created QPs may have implicit priority in NIC scheduling? So changing the order will balance the send behavior.

@LUIGINOVA5 commented on GitHub (May 9, 2025): > 1. Yes, recv queue can be removed. > 2. Changing the QP creation order is intended to make the NIC's send behavior more balanced. Thanks for reply! Regarding answer 2, earlier-created QPs may have implicit priority in NIC scheduling? So changing the order will balance the send behavior.

bsakaguchi commented

2026-10-26 09:54:56 +00:00

Author

Owner

@sphish commented on GitHub (May 9, 2025):

According to our experiments, the NIC schedules send operations in a round-robin manner based on the QPN. QPs created earlier have smaller QPNs and are scheduled first.

@sphish commented on GitHub (May 9, 2025): According to our experiments, the NIC schedules send operations in a round-robin manner based on the QPN. QPs created earlier have smaller QPNs and are scheduled first.

bsakaguchi commented

2026-10-26 09:54:56 +00:00

Author

Owner

@LUIGINOVA5 commented on GitHub (May 9, 2025):

According to our experiments, the NIC schedules send operations in a round-robin manner based on the QPN. QPs created earlier have smaller QPNs and are scheduled first.

Got it! Thanks!

@LUIGINOVA5 commented on GitHub (May 9, 2025): > According to our experiments, the NIC schedules send operations in a round-robin manner based on the QPN. QPs created earlier have smaller QPNs and are scheduled first. Got it! Thanks!

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepEP#129

No description provided.

Rows
Columns