Optimizing Distributed ML Communication with Fused Computation-Collective Operations

In order to satisfy their ever increasing capacity and compute requirements, machine learning models are distributed across multiple nodes using numerous parallelism strategies. As a result, collective communications are often on the critical path, and hiding their latency by overlapping kernel-granular communication and computation is difficult due to the absence of independent computation.

In this work, we propose fusing computation with dependent collective communication by leveraging GPUs' massive parallelism and GPU-initiated communication. We have developed selfcontained GPU kernels where threadblocks/workgroups (WGs) immediately communicate their results to remote GPUs when they complete their computation. Meanwhile, other WGs within the same kernel perform overlapping computation, maintaining high ALU utilization. Such fine-grain overlapping provides the additional benefit that peak network bandwidth demand is reduced and communication is spread across the entire lifetime of application rather than only at kernel boundaries. Furthermore, we propose zero-copy optimizations for scale-up communication where the data computed by one GPU is directly communicated to peer GPUs, eliminating intermediate stores and buffering.

We demonstrate our approach by creating three prototype fused operators (embedding + All-to-All, GEMV + AllReduce, and GEMM + All-to-All) to address the pervasive communication overheads observed in deep learning recommendation models (DLRM), Transformers and Mixture of Experts (MoE) model architectures. In order to demonstrate that our approach can be integrated into ML frameworks for wide adoption in production environments, we expose our fused operators as new PyTorch operators as well as extend the Triton framework to enable them. Our evaluations show that our approach can effectively overlap communication with computations, subsequently reducing their combined execution time than the current collective library-based approaches. Our scale-up GEMV + AllReduce and GEMM + Allto-All implementations achieve 13% (up to 22%) and 12% (up to 20%) lower execution time, while our fused embedding + All-to-All reduces execution time by 20% and 31% for intra-node and inter-node configurations. Large scale-out simulations indicate that our approach reduces DLRM execution time by 21% for 128 node system.