Improving the Performance of Distributed MXNet with RDMA

摘要

As one of the most influential deep learning frameworks, MXNet has achieved excellent performance and many breakthroughs in academic and industrial fields for various machine learning situations. The initial implementation of MXNet uses proxy-socket interface, which delivers suboptimal performance in distributed environment. In a massive parallel training task, parameters are updated frequently during each training loop, in which case network performance becomes the main factor of overall performance. Over the past decade, high performance interconnects have employed remote direct memory access (RDMA) technology to provide excellent performance for numerous scientific domains. In this paper, we describe an efficient design that extends the open-source MXNet to make it RDMA capable via RDMA-based parameter server interfaces. With modest optimizations towards memory usage and transmission overhead, RDMA-based MXNet achieves great performance improvement over the original software. Our experiments reveal that, for the communication subsystem of MXNet, the new design achieves 16x speedup (up to 21x at peak) over 1 Gigabit Ethernet (1GigE). For the two training cases on MXNet, the optimized implementation gains 5x and 9x speedup, respectively. Compared to experiments on the IP-over-InfiniBand (IPoIB) protocol, it achieves nearly 30% performance improvement, as well as better scalability and alleviation of bottlenecks.