PhaseSync: Look-Ahead Optical Circuit Scheduling for Multi-Job Distributed Training
Abstract
AI training workloads generate structured, predictable traffic that alternates between computation and synchronized collective operations, yet existing network schedulers do not fully exploit this structure for scheduling. Reconfigurable networks can exploit this structure by dynamically adapting connectivity to communication demand, enabling more efficient use of network capacity than static electrically switched fabrics. However, the bursty nature of this traffic can still create transient congestion, as synchronized traffic may be injected faster than transport protocols can react, leading to job slowdown. This problem is exacerbated in multi-job clusters, where communication bursts from concurrent jobs overlap and contend for shared network resources. We argue that this workload structure can be exploited to manage contention proactively, and propose PhaseSync, an event-driven Integer Linear Programming (ILP)-based scheduler that prioritizes competing collectives according to their estimated impact on job progress and allocates optical circuits accordingly, with the goal of reducing contention and overall makespan while preserving fairness and efficient network utilization.