CrossoverScheduler: Overlapping Multiple Distributed Training Applications in a Crossover Manner

TL;DR

CrossoverScheduler is a novel algorithm that enables multiple distributed deep learning tasks to share GPU resources through pipelined communication and computation, significantly improving training speed without affecting accuracy.

Contribution

It introduces Crossover Synchronization, allowing multiple applications to time-share GPUs, and demonstrates its effectiveness with a prototype integrated into Horovod.

Findings

Achieves 20% speedup on ImageNet classification

Enables overlapping multiple training applications without accuracy loss

Improves distributed training throughput significantly

Abstract

Distributed deep learning workloads include throughput-intensive training tasks on the GPU clusters, where the Distributed Stochastic Gradient Descent (SGD) incurs significant communication delays after backward propagation, forces workers to wait for the gradient synchronization via a centralized parameter server or directly in decentralized workers. We present CrossoverScheduler, an algorithm that enables communication cycles of a distributed training application to be filled by other applications through pipelining communication and computation. With CrossoverScheduler, the running performance of distributed training can be significantly improved without sacrificing convergence rate and network accuracy. We achieve so by introducing Crossover Synchronization which allows multiple distributed deep learning applications to time-share the same GPU alternately. The prototype of CrossoverScheduler is built and integrated with Horovod. Experiments on a variety of distributed tasks show that CrossoverScheduler achieves 20% \times speedup for image classification tasks on ImageNet dataset.