Consistent Lock-free Parallel Stochastic Gradient Descent for Fast and Stable Convergence

Abstract: Stochastic gradient descent (SGD) is an essential element in Machine Learning (ML) algorithms. Asynchronous parallel shared-memory SGD (AsyncSGD), including synchronization-free algorithms, e.g. HOGWILD!, have received interest in certain contexts, due to reduced overhead compared to synchronous parallelization. Despite that they induce staleness and inconsistency, they have shown speedup for problems satisfying smooth, strongly convex targets, and gradient sparsity. Recent works take important steps towards understanding the potential of parallel SGD for problems not conforming to these strong assumptions, in particular for deep learning (DL). There is however a gap in current literature in understanding when AsyncSGD algorithms are useful in practice, and in particular how mechanisms for synchronization and consistency play a role. We focus on the impact of consistency-preserving non-blocking synchronization in SGD convergence, and in sensitivity to hyper-parameter tuning. We propose Leashed-SGD, an extensible algorithmic framework of consistency-preserving implementations of AsyncSGD, employing lock-free synchronization, effectively balancing throughput and latency. We argue analytically about the dynamics of the algorithms, memory consumption, the threads' progress over time, and the expected contention. We provide a comprehensive empirical evaluation, validating the analytical claims, benchmarking the proposed Leashed-SGD framework, and comparing to baselines for training multilayer perceptrons (MLP) and convolutional neural networks (CNN). We observe the crucial impact of contention, staleness and consistency and show how Leashed-SGD provides significant improvements in stability as well as wall-clock time to convergence (from 20-80% up to 4x improvements) compared to the standard lock-based AsyncSGD algorithm and HOGWILD!, while reducing the overall memory footprint.