Abstract

The H3C AD-DC path navigation solution is an innovative traffic load balancing optimization solution for intelligent computing networks. By sensing network topology, server topology, and communication  traffic characteristics, H3C AD-DC actively plans service traffic paths from a global perspective, on both the server side and the network side. It can dynamically adjust the scheduling policy based on the actual network operation to avoid network congestion, achieve optimal traffic load sharing, and significantly improve model training efficiency.

The Tolly test used 64 NVIDIA GPUs to verify the performance of the path navigation solution and the traditional ECMP solution in the same service scenario. The test results show that, in the same service scenario, the path navigation solution has a significant improvement in bandwidth performance (busbw) compared to the traditional ECMP solution. Additionally, AD-DC can provide comprehensive operations support throughout the training process by monitoring key metrics of the network. By associating and comparing data to infer and analyze, it can detect changes in key metrics before and after faults, clearly identify the root cause and solution, significantly enhancing the stability of model training.