AET-SGD: Asynchronous Event-triggered Stochastic Gradient Descent

TL;DR

AET-SGD is an asynchronous event-triggered stochastic gradient descent framework that significantly reduces communication costs in distributed learning while maintaining good convergence and robustness against delays.

Contribution

It introduces a novel linear increasing sample size event-triggered threshold for asynchronous SGD, effectively reducing communication and mitigating delay impacts.

Findings

Achieves 44x to 120x reduction in communication cost.

Maintains convergence performance despite large delays.

Validated on multiple datasets including MNIST and CIFAR10.

Abstract

Communication cost is the main bottleneck for the design of effective distributed learning algorithms. Recently, event-triggered techniques have been proposed to reduce the exchanged information among compute nodes and thus alleviate the communication cost. However, most existing event-triggered approaches only consider heuristic event-triggered thresholds. They also ignore the impact of computation and network delay, which play an important role on the training performance. In this paper, we propose an Asynchronous Event-triggered Stochastic Gradient Descent (SGD) framework, called AET-SGD, to i) reduce the communication cost among the compute nodes, and ii) mitigate the impact of the delay. Compared with baseline event-triggered methods, AET-SGD employs a linear increasing sample size event-triggered threshold, and can significantly reduce the communication cost while keeping good convergence performance. We implement AET-SGD and evaluate its performance on multiple representative data sets, including MNIST, FashionMNIST, KMNIST and CIFAR10. The experimental results validate the correctness of the design and show a significant communication cost reduction from 44x to 120x, compared to the state of the art. Our results also show that AET-SGD can resist large delay from the straggler nodes while obtaining a decent performance and a desired speedup ratio.