Exploiting Multicast for Accelerating Collective Communication

TL;DR

This paper introduces MultiWrite, a multicast-based method that reduces collective communication latency in AI workloads by eliminating redundant data transmissions, leading to significant latency improvements.

Contribution

MultiWrite is a novel many-to-many transmission semantic that overcomes traditional multicast limitations for AI workloads, improving latency and network efficiency.

Findings

Achieves up to 33% latency reduction on Ascend NPUs.

Reduces network congestion by eliminating redundant data packets.

Demonstrates effectiveness in large-scale AI model training and inference.

Abstract

Reducing collective communication latency is a critical goal for large model training and inference in both academia and industry. Many-to-many communications, such as AllGather and AlltoAll (dispatch), are core components of modern parallelization strategies. State-of-the-art implementations of these communications rely on unicast-based writes and transmit duplicate copies of the same data across physical links for multiple receivers. This redundant transmission congests network bottlenecks and degrades end-to-end latency. We present MultiWrite, a novel many-to-many transmission semantic that eliminates redundant packets to directly reduce operator latency. MultiWrite adopts multicast principles while addressing critical limitations of traditional multicast for AI workloads. These limitations include heavy management plane overhead and ecosystem compatibility issues. We implement MultiWrite on Ascend NPUs. Long-term stress tests demonstrate that our MultiWrite-based operators achieve up to 33% latency reduction on commercially deployed devices.