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+.

tests/test_internode.py

@@ -218,17 +218,19 @@ def test_main(num_sms: int, local_rank: int, num_local_ranks: int, num_ranks: in
 # noinspection PyUnboundLocalVariable
 def test_loop(local_rank: int, num_local_ranks: int):
     num_nodes = int(os.getenv('WORLD_SIZE', 1))
+    num_sms = 24
+    qp_num = num_sms // 2
     rank, num_ranks, group = init_dist(local_rank, num_local_ranks)
     test_ll_compatibility = False
     if test_ll_compatibility:
         ll_num_tokens, ll_hidden, ll_num_experts, ll_num_topk = 16, 5120, 256, 9
 
     buffer = deep_ep.Buffer(group, int(1e9), int(1e9), low_latency_mode=test_ll_compatibility,
-                            num_qps_per_rank=(ll_num_experts // num_ranks if test_ll_compatibility else 1))
+                            num_qps_per_rank=(ll_num_experts // num_ranks if test_ll_compatibility else qp_num))
     assert num_local_ranks == 8 and num_ranks > 8
     torch.manual_seed(rank)
 
-    for i in (24, ):
+    for i in (num_sms, ):
         test_main(i, local_rank, num_local_ranks, num_ranks, num_nodes, rank, buffer, group)
         if local_rank == 0:
             print('', flush=True)