Nexus: Bringing Efficient and Scalable Training to Deep Learning Frameworks

摘要

Demand is mounting in the industry for scalable GPU-based deep learning systems. Unfortunately, existing training applications built atop popular deep learning frameworks, including Caffe, Theano, and Torch, etc, are incapable of conducting distributed GPU training over large-scale clusters. To remedy such a situation, this paper presents Nexus, a platform that allows existing deep learning frameworks to easily scale out to multiple machines without sacrificing model accuracy. Nexus leverages recently proposed distributed parameter management architecture to orchestrate distributed training by a large number of learners spread across the cluster. Through characterizing the run-time behavior of existing single-node based applications, Nexus is equipped with a suite of optimization schemes, including hierarchical and hybrid parameter aggregation, enhanced network and computation layer, and quality-guided communication adjustment, etc, to strengthen the communication channels and resource utilization. Empirical evaluations with a diverse set of deep learning applications demonstrate that Nexus is easy to integrate and can deliver efficient distributed training services to major deep learning frameworks. In addition, Nexus's optimization schemes are highly effective to shorten the training time with targeted accuracy bounds.