Anytime Stochastic Gradient Descent: A Time to Hear from all the Workers
This addresses the straggler problem in parallel SGD for machine learning practitioners, offering a significant but incremental improvement over prior synchronization approaches.
The paper tackles the problem of slow workers (stragglers) stalling parallel stochastic gradient descent (SGD) by proposing a method that fixes computation time and combines all workers' contributions, resulting in an improvement of several factors of magnitude compared to existing methods.
In this paper, we focus on approaches to parallelizing stochastic gradient descent (SGD) wherein data is farmed out to a set of workers, the results of which, after a number of updates, are then combined at a central master node. Although such synchronized SGD approaches parallelize well in idealized computing environments, they often fail to realize their promised computational acceleration in practical settings. One cause is slow workers, termed stragglers, who can cause the fusion step at the master node to stall, which greatly slowing convergence. In many straggler mitigation approaches work completed by these nodes, while only partial, is discarded completely. In this paper, we propose an approach to parallelizing synchronous SGD that exploits the work completed by all workers. The central idea is to fix the computation time of each worker and then to combine distinct contributions of all workers. We provide a convergence analysis and optimize the combination function. Our numerical results demonstrate an improvement of several factors of magnitude in comparison to existing methods.