[GH-ISSUE #608] Hybrid_EP Abnormal performance on B300 (2-node dispatch ~40GB/s) reveals hardcoded DEF_IB_TC; request env-based configurability (align with NVSHMEM) #368

New issue

Open

opened 2026-10-26 09:59:14 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:59:14 +00:00

Owner

Originally created by @kzlxd on GitHub (Apr 24, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/608

Description:

In the current Hybrid EP implementation, the InfiniBand Traffic Class is hardcoded as:

constexpr int32_t DEF_IB_TC = 0;

This makes the IB traffic class fixed at compile time and prevents users from tuning QoS / traffic prioritization dynamically based on deployment environments.

In contrast, NVSHMEM provides environment variable support (e.g., NVSHMEM_IB_TC) to configure the traffic class at runtime, which is very useful for performance tuning and congestion control in RDMA-based communication.

Problem:

DEF_IB_TC cannot be adjusted without recompilation
No flexibility for different cluster QoS configurations
Inconsistent behavior compared to NVSHMEM and other RDMA-based stacks

Suggestion:
Introduce an environment variable (e.g., HybridEP_IB_TC) to override the default value. For example:

int32_t get_ib_tc() {
    const char* env = std::getenv("HybridEP_IB_TC");
    if (env != nullptr) {
        return std::atoi(env);
    }
    return DEF_IB_TC;
}

Then use get_ib_tc() instead of the compile-time constant.

Benefits:

Enables runtime tuning without recompilation
Aligns with NVSHMEM and common RDMA practices
Improves flexibility for performance optimization under different network conditions

Would it be possible to add this support for better configurability?

Originally created by @kzlxd on GitHub (Apr 24, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/608 **Description:** In the current Hybrid EP implementation, the InfiniBand Traffic Class is hardcoded as: ```cpp constexpr int32_t DEF_IB_TC = 0; ``` This makes the IB traffic class fixed at compile time and prevents users from tuning QoS / traffic prioritization dynamically based on deployment environments. In contrast, NVSHMEM provides environment variable support (e.g., `NVSHMEM_IB_TC`) to configure the traffic class at runtime, which is very useful for performance tuning and congestion control in RDMA-based communication. **Problem:** * `DEF_IB_TC` cannot be adjusted without recompilation * No flexibility for different cluster QoS configurations * Inconsistent behavior compared to NVSHMEM and other RDMA-based stacks **Suggestion:** Introduce an environment variable (e.g., `HybridEP_IB_TC`) to override the default value. For example: ```cpp int32_t get_ib_tc() { const char* env = std::getenv("HybridEP_IB_TC"); if (env != nullptr) { return std::atoi(env); } return DEF_IB_TC; } ``` Then use `get_ib_tc()` instead of the compile-time constant. **Benefits:** * Enables runtime tuning without recompilation * Aligns with NVSHMEM and common RDMA practices * Improves flexibility for performance optimization under different network conditions Would it be possible to add this support for better configurability?

bsakaguchi commented

2026-10-26 09:59:16 +00:00

Author

Owner

@Autumn1998 commented on GitHub (May 9, 2026):

Thanks! we will consider this in the next version.

@Autumn1998 commented on GitHub (May 9, 2026): Thanks! we will consider this in the next version.

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#368

No description provided.

Rows
Columns