Local SGD Converges Fast and Communicates Little
read the original abstract
Mini-batch stochastic gradient descent (SGD) is state of the art in large scale distributed training. The scheme can reach a linear speedup with respect to the number of workers, but this is rarely seen in practice as the scheme often suffers from large network delays and bandwidth limits. To overcome this communication bottleneck recent works propose to reduce the communication frequency. An algorithm of this type is local SGD that runs SGD independently in parallel on different workers and averages the sequences only once in a while. This scheme shows promising results in practice, but eluded thorough theoretical analysis. We prove concise convergence rates for local SGD on convex problems and show that it converges at the same rate as mini-batch SGD in terms of number of evaluated gradients, that is, the scheme achieves linear speedup in the number of workers and mini-batch size. The number of communication rounds can be reduced up to a factor of T^{1/2}---where T denotes the number of total steps---compared to mini-batch SGD. This also holds for asynchronous implementations. Local SGD can also be used for large scale training of deep learning models. The results shown here aim serving as a guideline to further explore the theoretical and practical aspects of local SGD in these applications.
This paper has not been read by Pith yet.
Forward citations
Cited by 13 Pith papers
-
Sensing-Native Over-the-Air Federated Learning
Proposes a zero-overhead sensing-native over-the-air FL system that uses gradient autocorrelation for distance estimation and gradient statistics for cooperative localization, with a statistics-aware power and beamfor...
-
Demystifying Pipeline Parallelism: First Theory for PipeDream
Introduces Randomized PipeDream abstraction yielding first nonconvex convergence bound for PipeDream and proves delay scales as S squared for S stages.
-
Gaussian Approximation and Multiplier Bootstrap for Federated Linear Stochastic Approximation
Establishes non-asymptotic Gaussian approximation bounds for federated LSA with explicit communication-heterogeneity trade-offs and introduces an online multiplier bootstrap for last-iterate inference with validity gu...
-
Distributionally Robust Multi-Task Reinforcement Learning via Adaptive Task Sampling
DRATS derives a minimax objective from a feasibility formulation of MTRL to adaptively sample tasks with the largest return gaps, leading to better worst-task performance on MetaWorld benchmarks.
-
Unifying Local Communications and Local Updates for LLM Pretraining
GASLoC generalizes communication acceleration to the outer optimizer to enable gossip-based decentralized LLM pretraining that supports adaptive optimizers, local steps, and outperforms prior decentralized methods on ...
-
Boosting Multimodal Federated Learning via Chained Modality Optimization
FedMChain improves multimodal federated learning by chaining modality-wise optimization phases with error-compensated regularization and sparse sign-guided aggregation to mitigate modality competition and cut communic...
-
Synchronous and Asynchronous Parallelism Approaches for Generalized Canonical Polyadic Tensor Decomposition with GenTen
Presents new synchronous and asynchronous parallel approaches for GCP tensor decomposition and evaluates computational cost and accuracy on synthetic and real-world datasets.
-
Statistical Limits and Efficient Algorithms for Differentially Private Federated Learning
Introduces FedHybrid and FedNewton for DP federated M-estimation, with finite-sample MSE bounds, minimax lower bound, and evaluations on vision datasets.
-
Runtime-Orchestrated Second-Order Optimization for Scalable LLM Training
Asteria is a runtime system that enables second-order optimization for LLMs by dynamically distributing optimizer state across GPU, CPU, and NVMe while using asynchronous inverse-root computations and bounded-stalenes...
-
Response-Conditioned Parallel-to-Sequential Orchestration for Multi-Agent Systems
Nexa learns a response-conditioned policy that starts with parallel agent execution and adds at most one round of sequential message passing via a predicted sparse DAG, strictly subsuming pure parallel mode.
-
Rescaled Asynchronous SGD: Optimal Distributed Optimization under Data and System Heterogeneity
Rescaled ASGD recovers convergence to the true global objective by rescaling worker stepsizes proportional to computation times, matching the known time lower bound in the leading term under non-convex smoothness and ...
-
A Note on Stability for Orthogonalized Matrix Momentum with Client Sampling
Derives finite-round upper-tail guarantee on population-empirical gap for client-sampled orthogonalized matrix momentum under heterogeneous data, with Lipschitz condition on the orthogonalizer.
-
From Data Heterogeneity to Convergence: A Data-Centric Review of Federated Learning
A data-centric survey of federated learning that ranks non-IID data traits by influence on convergence, links splitting protocols to real phenomena, and examines data-related defenses under clean and adversarial conditions.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.