HiPS: Hierarchical parameter synchronization in large-scale distributed machine learning

Abstract

In large-scale distributed machine learning (DML) system, parameter (gradient) synchronization among machines plays an important role in improving the DML performance. State-of-the-art DML synchronization algorithms, either the parameter server (PS) based algorithm or the ring allreduce algorithm, work in a flat way and suffer when the network size is large. In this work, we propose HiPS, a hierarchical parameter (gradient) synchronization framework in large-scale DML. In HiPS, server-centric network topology is used to better embrace RDMA/RoCE transport between machines, and the parameters (gradients) are synchronized in a hierarchical and hybrid way. Our evaluation in BCube and Torus network demonstrates that HiPS can better match server-centric networks. Compared with the flat algorithms (PS-based and ring-based), HiPS reduces the synchronization time by 73% and 75% respectively.