Resilient AI Supercomputer Networking using MRC and SRv6
For organizations running large-scale AI training, this work provides a practical networking solution that reduces tail latency and improves resilience to failures, enabling more efficient training of frontier models.
The paper addresses tail latency in large-scale synchronous pretraining by introducing MRC, a new RDMA transport protocol that sprays across multiple paths and load-balances, combined with multi-plane Clos topologies and SRv6 static source-routing. In production at OpenAI and Microsoft, this approach enables training clusters over 100K GPUs and allows training jobs to survive network failures that previously caused interruptions.
Tail latency dominates the performance of synchronous pretraining jobs when running at very large scales. We describe a three-pronged approach: (1) a new RDMA-based transport protocol, MRC, sprays across many paths and actively load-balances between them, eliminating the issue of flow collisions (2) the use of multi-plane Clos topologies to get the benefits of high switch radix and redundancy, allowing training clusters well over 100K GPUs to be built as two-tier topologies while increasing physical redundancy, and (3) the use of static source-routing using SRv6 to allow MRC the freedom to bypass failures by itself. We describe our experiences running MRC and static SRv6 routing in production in OpenAI and Microsoft's largest training clusters, where it has been used to train the latest frontier models. We demonstrate how MRC allows AI training jobs to ride out many network failures that previously would have interrupted training.