mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #60] RDMA bandwidth is not normal when run test_internode.py. #57
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepEP#57
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @MinhuiWan on GitHub (Mar 9, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/60
My test environment is 2 nodes, 8 H100 per node, 4 IB per node。 the test result is
[tuning] Best dispatch (FP8): SMs 24, NVL chunk 20, RDMA chunk 32: 4.47 GB/s (RDMA), 9.00 GB/s (NVL)
[tuning] Best dispatch (BF16): SMs 24, NVL chunk 24, RDMA chunk 32: 4.72 GB/s (RDMA), 9.51 GB/s (NVL)
[tuning] Best combine: SMs 24, NVL chunk 4, RDMA chunk 32: 4.71 GB/s (RDMA), 9.50 GB/s (NVL)
But in readme.md, how many ib cards per node?
During test, arethere any environment variables that need to be set?
@Huixxi commented on GitHub (Mar 9, 2025):
Is it pure roce rdma or through IB?
@sphish commented on GitHub (Mar 10, 2025):
Our test results were obtained using 8 H800 and 8 IB nodes. We haven't encountered the issue you're experiencing. Perhaps the 4 IB setup requires some additional configurations. There is someone who encountered the same issue https://github.com/deepseek-ai/DeepEP/issues/38#issuecomment-2703724191 and found a solution, so you can pay attention to his response.
@sphish commented on GitHub (Mar 10, 2025):
Can you run
nvidia-smi topo -mpandibv_devinfoand share the results?@MinhuiWan commented on GitHub (Mar 10, 2025):
During test, i set the environment variable with this command:
export NVSHMEM_ENABLE_NIC_PE_MAPPING=1
export NVSHMEM_HCA_LIST=mlx5_1:1,mlx5_7:1,mlx5_9:1,mlx5_13:1
This configuration makes NVSHMEM use the specified IB ports.
@MinhuiWan commented on GitHub (Mar 10, 2025):
I can't understand what‘s relationship of rdma ans IB。During test, i set the environment variable with this command:
export NVSHMEM_ENABLE_NIC_PE_MAPPING=1
export NVSHMEM_HCA_LIST=mlx5_1:1,mlx5_7:1,mlx5_9:1,mlx5_13:1
This configuration makes NVSHMEM use the specified IB ports.
@MinhuiWan commented on GitHub (Mar 10, 2025):
@MinhuiWan commented on GitHub (Mar 18, 2025):
Ensure 8IB per node, and use nccl_test to test the throughout of alltoall, when environment is normal,run test internode.py,the RDMA bandwidth is normal。
@CSEEduanyu commented on GitHub (Apr 23, 2025):
Hello, could you please explain in detail how Ensure 8IB per node operates?