Asynchronous SGD Beats Minibatch SGD Under Arbitrary Delays
This work addresses performance issues in distributed machine learning for practitioners using parallel computing, though it appears incremental as it improves upon existing analyses.
The paper tackled the problem of asynchronous SGD's performance degradation under large delays by proving better guarantees that depend on the number of parallel devices rather than delays, showing it outperforms synchronous minibatch SGD in certain settings.
The existing analysis of asynchronous stochastic gradient descent (SGD) degrades dramatically when any delay is large, giving the impression that performance depends primarily on the delay. On the contrary, we prove much better guarantees for the same asynchronous SGD algorithm regardless of the delays in the gradients, depending instead just on the number of parallel devices used to implement the algorithm. Our guarantees are strictly better than the existing analyses, and we also argue that asynchronous SGD outperforms synchronous minibatch SGD in the settings we consider. For our analysis, we introduce a novel recursion based on "virtual iterates" and delay-adaptive stepsizes, which allow us to derive state-of-the-art guarantees for both convex and non-convex objectives.