Straggler-Aware Distributed Learning: Communication–Computation Latency Trade-Off

Abstract

When gradient descent (GD) is scaled to many parallel workers for large-scale machine learning applications, its per-iteration computation time is limited by straggling workers. Straggling workers can be tolerated by assigning redundant computations and/or coding across data and computations, but in most existing schemes, each non-straggling worker transmits one message per iteration to the parameter server (PS) after completing all its computations. Imposing such a limitation results in two drawbacks: over-computation due to inaccurate prediction of the straggling behavior, and under-utilization due to discarding partial computations carried out by stragglers. To overcome these drawbacks, we consider multi-message communication (MMC) by allowing multiple computations to be conveyed from each worker per iteration, and propose novel straggler avoidance techniques for both coded computation and coded communication with MMC. We analyze how the proposed designs can be employed efficiently to seek a balance between the computation and communication latency. Furthermore, we identify the advantages and disadvantages of these designs in different settings through extensive simulations, both model-based and real implementation on Amazon EC2 servers, and demonstrate that proposed schemes with MMC can help improve upon existing straggler avoidance schemes.