STSyn: Speeding Up Local SGD With Straggler-Tolerant Synchronization

Synchronous local stochastic gradient descent (local SGD) suffers from some workers being idle and random delays due to slow and straggling workers, as it waits for the workers to complete the same amount of local updates. To address this issue, a novel local SGD strategy called STSyn is proposed in this paper. The key point is to wait for the $K$ fastest workers while keeping all the workers computing continually at each synchronization round, and making full use of any effective (completed) local update of each worker regardless of stragglers. To evaluate the performance of STSyn, an analysis of the average wall-clock time, average number of local updates, and average number of uploading workers per round is provided. The convergence of STSyn is also rigorously established even when the objective function is nonconvex for both homogeneous and heterogeneous data distributions. Experimental results highlight the superiority of STSyn over state-of-the-art schemes, thanks to its straggler-tolerant technique and the inclusion of additional effective local updates at each worker. Furthermore, the impact of system parameters is investigated. By waiting for faster workers and allowing heterogeneous synchronization with different numbers of local updates across workers, STSyn provides substantial improvements both in time and communication efficiency.