Efficient All-to-All Collective Communication Schedules for Direct-Connect Topologies

The all-to-all collective communications primitive is widely used in machine learning (ML) and high performance computing (HPC) workloads, and optimizing its performance is of interest to both ML and HPC communities. All-to-all is a particularly challenging workload that can severely strain the underlying interconnect bandwidth at scale. This is mainly because of the quadratic scaling in the number of messages that must be simultaneously serviced combined with large message sizes. This paper takes a holistic approach to optimize the performance of all-to-all collective communications on supercomputer-scale direct-connect interconnects. We address several algorithmic and practical challenges in developing efficient and bandwidth-optimal all-to-all schedules for any topology, lowering the schedules to various backends and fabrics that may or may not expose additional forwarding bandwidth, establishing an upper bound on all-to-all throughput, and exploring novel topologies that deliver near-optimal all-to-all performance.