DSAG: A mixed synchronous-asynchronous iterative method for straggler-resilient learning

TL;DR

This paper introduces DSAG, a novel mixed synchronous-asynchronous iterative method for resilient distributed learning that effectively handles persistent stragglers, demonstrating significant speed improvements over existing methods in practical cloud computing scenarios.

Contribution

The paper proposes DSAG, a new iterative optimization algorithm combining synchronous and asynchronous updates, with a latency model and load-balancing strategy tailored for persistent stragglers.

Findings

DSAG is up to 50% faster than SAG.

DSAG is more than twice as fast as coded computing methods.

Effective in large-scale genomics and logistic regression tasks.

Abstract

We consider straggler-resilient learning. In many previous works, e.g., in the coded computing literature, straggling is modeled as random delays that are independent and identically distributed between workers. However, in many practical scenarios, a given worker may straggle over an extended period of time. We propose a latency model that captures this behavior and is substantiated by traces collected on Microsoft Azure, Amazon Web Services (AWS), and a small local cluster. Building on this model, we propose DSAG, a mixed synchronous-asynchronous iterative optimization method, based on the stochastic average gradient (SAG) method, that combines timely and stale results. We also propose a dynamic load-balancing strategy to further reduce the impact of straggling workers. We evaluate DSAG for principal component analysis, cast as a finite-sum optimization problem, of a large genomics dataset, and for logistic regression on a cluster composed of 100 workers on AWS, and find that DSAG is up to about 50% faster than SAG, and more than twice as fast as coded computing methods, for the particular scenario that we consider.