Consistent Lock-free Parallel Stochastic Gradient Descent for Fast and Stable Convergence
| Autor: | Karl Bäckström, Marina Papatriantafilou, Philippas Tsigas, Ivan Walulya |
|---|---|
| Rok vydání: | 2021 |
| Předmět: |
FOS: Computer and information sciences
Theoretical computer science Speedup Artificial neural network Computer science business.industry Deep learning Parallel algorithm Stability (learning theory) Perceptron Convolutional neural network Memory management Stochastic gradient descent Computer Science - Distributed Parallel and Cluster Computing Synchronization (computer science) Computer Science - Data Structures and Algorithms Overhead (computing) Data Structures and Algorithms (cs.DS) Artificial intelligence Distributed Parallel and Cluster Computing (cs.DC) Convex function business |
| Zdroj: | IPDPS |
| DOI: | 10.48550/arxiv.2102.09032 |
| Popis: | 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. Comment: 13 pages, 10 figures. Accepted in the 35th IEEE International Parallel & Distributed Processing Symposium |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|