FSP: towards flexible synchronous parallel framework for expectation-maximization based algorithms on cloud

Abstract

Myriad of parameter estimation algorithms can be performed by an Expectation-Maximization (EM) approach. Traditional synchronous frameworks can parallelize these EM algorithms on the cloud to accelerate computation while guaranteeing the convergence. However, expensive synchronization costs pose great challenges for efficiency. Asynchronous solutions have been recently designed to bypass high-cost synchronous barriers but at expense of potentially losing convergence guarantee. This paper first proposes a flexible synchronous parallel framework (FSP) that provides the capability of synchronous EM algorithms implementations, as well as significantly reduces the barrier cost. Under FSP, every distributed worker can immediately suspend local computation when necessary, to quickly synchronize with each other. That maximizes the time fast workers spend doing useful work, instead of waiting for slow, straggling workers. We then formally prove the algorithm convergence. Further, we analyze how to automatically identify a proper barrier interval to strike a nice balance between reduced synchronization costs and the convergence speed. Empirical results demonstrate that on a broad spectrum of real-world and synthetic datasets, FSP achieves as much as 3x speedup over the up-to-date synchronous solution.