Task allocation for decentralized training in heterogeneous environment

TL;DR

This paper proposes a static and self-adaptive data allocation algorithm for decentralized training in heterogeneous environments, significantly reducing training time and improving resource utilization.

Contribution

It introduces a novel self-adaptive data allocation method that enhances decentralized training efficiency in heterogeneous clusters, outperforming proportional allocation.

Findings

Self-adaptive allocation reduces training time by nearly 50%.

Algorithm improves resource utilization in heterogeneous environments.

Training time decreases with better overall cluster performance.

Abstract

The demand for large-scale deep learning is increasing, and distributed training is the current mainstream solution. Ring AllReduce is widely used as a data parallel decentralized algorithm. However, in a heterogeneous environment, each worker calculates the same amount of data, so that there is a lot of waiting time loss among different workers, which makes the algorithm unable to adapt well to heterogeneous clusters. Resources are not used as they should be. In this paper, we design an implementation of static allocation algorithm. The dataset is artificially allocated to each worker, and samples are drawn proportionally for training, thereby speeding up the training speed of the network in a heterogeneous environment. We verify the convergence and influence on training speed of the network model under this algorithm on one machine with multi-card and multi-machine with multi-card. On this basis of feasibility, we propose a self-adaptive allocation algorithm that allows each machine to find the data it needs to adapt to the current environment. The self-adaptive allocation algorithm can reduce the training time by nearly one-third to half compared to the same proportional allocation.In order to better show the applicability of the algorithm in heterogeneous clusters, We replace a poorly performing worker with a good performing worker or add a poorly performing worker to the heterogeneous cluster. Experimental results show that training time will decrease as the overall performance improves. Therefore, it means that resources are fully used. Further, this algorithm is not only suitable for straggler problems, but also for most heterogeneous situations. It can be used as a plug-in for AllReduce and its variant algorithms.