On Scheduling DML Jobs in All-Optical DCNs with In-Network Computing

Enabled by programmable data plane (PDP), innetwork computing (INC) can offload the computation phase of distributed machine learning (DML) training to accelerate the execution of parameter servers (PS’). Meanwhile, all-optical interconnect (AOI) can effectively improve the network throughput for DML training. This work studies how to schedule a batch of parallel DML jobs in an all-optical data center network (DCN) to fully explore the mutual benefits of INC and AOI. Specifically, for each DML job, we determine where to deploy its workers, whether and where to offload/place its PS, and how to plan the routing of data transfers between the workers and the PS, such that its job completion time (JCT) is minimized. We formulate a mixed integer linear programming (MILP) model to optimize the job scheduling for the cases with and without INC, and propose a heuristic to tackle the case with INC quickly. Extensive simulations confirm the effectiveness of our proposals.