Full-Stack Allreduce on Multi-Rail Networks

The high communication costs impede scalability in distributed systems. Multimodal models like Sora exacerbate this issue by requiring more resources than current networks can support. However, existing network architectures fail to address this gap. In this paper, we provide full-stack support for allreduce on multi-rail networks, aiming to overcome the scalability limitations of large-scale networks by facilitating collaborative data transfer across various networks. To achieve this, we propose the Nezha system, which integrates TCP, in-network computing protocol SHARP, and RDMA-based protocol GLEX. To maximize data transfer rates, Nezha incorporates a load balancing data allocation scheme based on cost feedback and combines exception handling to achieve reliable data transmission. Our experiments on a six-node cluster demonstrate that Nezha significantly enhances allreduce performance by 58% to 87% in homogeneous dual-rail configurations and offers considerable acceleration in heterogeneous settings, contingent on the performance variance among networks.