Publication | Closed Access
Can Decentralized Algorithms Outperform Centralized Algorithms? A Case Study for Decentralized Parallel Stochastic Gradient Descent
442
Citations
0
References
2017
Year
Cluster ComputingEngineeringMachine LearningDistributed AlgorithmsNetwork AnalysisDistributed Ai SystemCentralized PsgdDistributed Data AnalyticsDistributed Problem SolvingParallel ComputingDistributed ModelDecentralised SystemComputer EngineeringComputer ScienceDistributed LearningCentralized AlgorithmsEdge ComputingCloud ComputingFederated LearningCase StudyParallel ProgrammingD-psgd Algorithm
Distributed ML systems such as TensorFlow and CNTK are typically centralized, creating a communication bottleneck at the central node, and prior analyses of decentralized PSGD have not demonstrated any advantage over centralized PSGD under standard assumptions. The study investigates whether decentralized PSGD can outperform centralized PSGD in distributed stochastic gradient descent. We analyze a D‑PSGD algorithm theoretically, showing it matches C‑PSGD’s computational complexity while reducing communication load, and empirically validate this across CNTK, Torch, varied networks, and up to 112 GPUs. In low‑bandwidth or high‑latency settings, D‑PSGD can be up to ten times faster than optimized centralized PSGD.
Most distributed machine learning systems nowadays, including TensorFlow and CNTK, are built in a centralized fashion. One bottleneck of centralized algorithms lies on high communication cost on the central node. Motivated by this, we ask, can decentralized algorithms be faster than its centralized counterpart? Although decentralized PSGD (D-PSGD) algorithms have been studied by the control community, existing analysis and theory do not show any advantage over centralized PSGD (C-PSGD) algorithms, simply assuming the application scenario where only the decentralized network is available. In this paper, we study a D-PSGD algorithm and provide the first theoretical analysis that indicates a regime in which decentralized algorithms might outperform centralized algorithms for distributed stochastic gradient descent. This is because D-PSGD has comparable total computational complexities to C-PSGD but requires much less communication cost on the busiest node. We further conduct an empirical study to validate our theoretical analysis across multiple frameworks (CNTK and Torch), different network configurations, and computation platforms up to 112 GPUs. On network configurations with low bandwidth or high latency, D-PSGD can be up to one order of magnitude faster than its well-optimized centralized counterparts.