Accelerating Distributed Deep Learning using Multi-Path RDMA in Data Center Networks

Abstract

Data center networks (DCNs) have widely deployed RDMA to support data-intensive applications such as machine learning. While DCNs are designed with rich multi-path topology, current RDMA (hardware) technology does not support multi-path transport. In this paper we advance Maestro- a purely software-basedmulti-path RDMA solution - to effectively utilize the rich multi-path topology for load balancing and reliability. As a "middleware" operating at the user-space, Maestro is modulaR@and software-defined:Maestro decouples path selection and load balancing mechanisms from hardware features, and allows DCN operators and applications to make flexible decisions by employing the best mechanisms as needed. As such, Maestro can be readily deployed using existing RDMA hardware (NICs) to support distributed deep learning (DDL) applications. Our experiments show that Maestro is capable of fully utilizing multiple paths with negligible CPU overheads, thereby enhancing the performance of DDL applications.