Abstract

 We introduce Bayesian distributed stochastic gradient descent (BDSGD), a highthroughput algorithm for training deep neural networks on parallel computing clusters. This algorithm uses amortized inference in a deep generative model to perform joint posterior predictive inference of mini-batch gradient computation times in a compute cluster specific manner. Specifically, our algorithm mitigates the straggler effect in synchronous, gradient-based optimization by choosing an optimal cutoff beyond which mini-batch gradient messages from slow workers are ignored. The principle novel contribution and finding of this work goes beyond this by demonstrating that using the predicted run-times from a generative model of cluster worker performance improves over the static-cutoff prior art, leading to higher gradient computation throughput on large compute clusters. In our experiments we show that eagerly discarding the mini-batch gradient computations of stragglers not only increases throughput but sometimes also increases the overall rate of convergence as a function of wall-clock time by virtue of eliminating idleness.