In the Internode Normal Kernel, when using nvshmem ibrc for RDMA data transmission, a single QP is used for data transfer between two GPUs, which limits kernel performance in network card dual-port and RoCE network scenarios.

In our optimized Internode Normal Kernel, we implemented multiple QPs for data transmission between two GPUs, setting a different QP for each channel. Additionally, we modified the transmission method from IBRC to IBGAD.

Through these optimizations, the Internode Normal Kernel achieves optimal performance in both H800 and H20 environments, with RDMA transmission performance nearly reaching the physical network performance limit. Using the current default statistical method, in 4-node H800 and H20 environments, RDMA performance can reach 60GB/s+.

deep_ep/buffer.py

@@ -65,9 +65,10 @@ class Buffer:
 
         # Synchronize NVSHMEM unique IDs
         root_unique_id = None
+        internode_use_ibgda = True
         if self.runtime.get_num_rdma_ranks() > 1 or low_latency_mode:
             # Enable IBGDA for the low latency mode, which refers to "no package forwarding between NVLink and RDMA"
-            if low_latency_mode:
+            if low_latency_mode or internode_use_ibgda:
                 assert num_qps_per_rank > 0
                 os.environ['NVSHMEM_DISABLE_P2P'] = '1'
                 os.environ['NVSHMEM_IB_ENABLE_IBGDA'] = '1'