{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/asynchronous-decentralized-parallel","title":"Asynchronous Decentralized Parallel Stochastic Gradient Descent","arxiv_id":"1710.06952","date":"2017-10-18","proceeding":"ICML 2018 7","authors":["Xiangru Lian","Wei zhang","Ce Zhang","Ji Liu"],"abstract":"Most commonly used distributed machine learning systems are either\nsynchronous or centralized asynchronous. Synchronous algorithms like\nAllReduce-SGD perform poorly in a heterogeneous environment, while asynchronous\nalgorithms using a parameter server suffer from 1) communication bottleneck at\nparameter servers when workers are many, and 2) significantly worse convergence\nwhen the traffic to parameter server is congested. Can we design an algorithm\nthat is robust in a heterogeneous environment, while being communication\nefficient and maintaining the best-possible convergence rate? In this paper, we\npropose an asynchronous decentralized stochastic gradient decent algorithm\n(AD-PSGD) satisfying all above expectations. Our theoretical analysis shows\nAD-PSGD converges at the optimal $O(1/\\sqrt{K})$ rate as SGD and has linear\nspeedup w.r.t. number of workers. Empirically, AD-PSGD outperforms the best of\ndecentralized parallel SGD (D-PSGD), asynchronous parallel SGD (A-PSGD), and\nstandard data parallel SGD (AllReduce-SGD), often by orders of magnitude in a\nheterogeneous environment. When training ResNet-50 on ImageNet with up to 128\nGPUs, AD-PSGD converges (w.r.t epochs) similarly to the AllReduce-SGD, but each\nepoch can be up to 4-8X faster than its synchronous counterparts in a\nnetwork-sharing HPC environment. To the best of our knowledge, AD-PSGD is the\nfirst asynchronous algorithm that achieves a similar epoch-wise convergence\nrate as AllReduce-SGD, at an over 100-GPU scale.","url_abs":"http://arxiv.org/abs/1710.06952v3","url_pdf":"http://arxiv.org/pdf/1710.06952v3.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"asynchronous-decentralized-parallel","repo_url":"https://github.com/VhalPurohit/290s","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"asynchronous-decentralized-parallel","repo_url":"https://github.com/adelnabli/acid","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}},{"paper_slug":"asynchronous-decentralized-parallel","repo_url":"https://github.com/facebookresearch/stochastic_gradient_push","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":null,"task_name":"GPU"}],"methods":[{"method_slug":"sgd","method_name":"SGD"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1710.06952","atlas_url":"https://app.syntology.ai/?focus=1710.06952","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}