Slow and Stale Gradients Can Win the Race: Error-Runtime Trade-offs in Distributed SGD
This work addresses performance bottlenecks in distributed machine learning training, offering incremental improvements for practitioners dealing with stragglers and staleness.
The paper tackles the trade-off between straggler delays and gradient staleness in distributed SGD, presenting a theoretical analysis that shows asynchronous methods can achieve better error-runtime trade-offs by balancing these factors, with results including a new convergence analysis and learning rate schedule.
Distributed Stochastic Gradient Descent (SGD) when run in a synchronous manner, suffers from delays in waiting for the slowest learners (stragglers). Asynchronous methods can alleviate stragglers, but cause gradient staleness that can adversely affect convergence. In this work we present a novel theoretical characterization of the speed-up offered by asynchronous methods by analyzing the trade-off between the error in the trained model and the actual training runtime (wallclock time). The novelty in our work is that our runtime analysis considers random straggler delays, which helps us design and compare distributed SGD algorithms that strike a balance between stragglers and staleness. We also present a new convergence analysis of asynchronous SGD variants without bounded or exponential delay assumptions, and a novel learning rate schedule to compensate for gradient staleness.