{"id":455,"date":"2025-08-14T21:41:21","date_gmt":"2025-08-14T13:41:21","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=455"},"modified":"2025-11-09T04:37:00","modified_gmt":"2025-11-08T20:37:00","slug":"cassini-network-aware-job-scheduling-in-machine-learning-clusters","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=455","title":{"rendered":"CASSINI: Network-Aware Job Scheduling in Machine Learning Clusters"},"content":{"rendered":"\n

\u4f5c\u8005:<\/strong>Sudarsanan Rajasekaran<\/a> (1), Manya Ghobadi<\/a> (1), Aditya Akella<\/a> (2) ((1) Massachusetts Institute of Technology, (2) UT Austin)<\/p>\n\n\n\n

\u8bba\u6587\u6458\u8981\u539f\u6587\uff1a<\/strong>We present CASSINI, a network-aware job scheduler for machine learning (ML) clusters<\/a>. CASSINI introduces a novel geometric abstraction to consider the communication pattern of different jobs while placing them on network links. To do so, CASSINI uses an affinity graph that finds a series of time-shift values to adjust the communication phases of a subset of jobs such that the communication patterns of jobs sharing the same network link are interleaved with each other. Experiments with 13 common ML models on a 24-server testbed demonstrate that compared to the state-of-the-ar<\/p>\n\n\n\n

t ML schedulers, CASSINI improves the average and tail completion time of jobs by up to 1.6\u00d7 and 2.5\u00d7,respectively. Moreover, we show that CASSINI reduces the number of ECN marked packets in the cluster by up to 33\u00d7.<\/p>\n\n\n\n

\"CASSINI:<\/figure>\n\n\n\n

\u7814\u7a76\u80cc\u666f<\/strong><\/h2>\n\n\n\n

\u8bad\u7ec3\u5927\u89c4\u6a21\u6a21\u578b\u9700\u8981\u4e00\u4e2a\u9ad8\u6548\u7684GPU\u96c6\u7fa4\u3002<\/p>\n\n\n\n

\u76f8\u5173\u7814\u7a76\u8868\u660e\u968f\u7740GPU\u6570\u91cf\u7684\u589e\u52a0\uff0c\u5206\u5e03\u5f0f\u8bad\u7ec3\u7684\u901a\u4fe1\u5f00\u9500\u5360\u603b\u65f6\u95f4\u5f00\u9500\u7684\u4e00\u5927\u90e8\u5206\u3002<\/p>\n\n\n\n

\u73b0\u6709\u6700\u5148\u8fdb\u7684\uff08state-of-the-art, SOTA\uff09\u6df1\u5ea6\u5b66\u4e60\u8c03\u5ea6\u5668\uff08Machine Learning Scheduler\uff09\u5728\u5206\u914d\u4efb\u52a1\u65f6\u5ffd\u7565\u4e86\u5206\u5e03\u5f0f\u8bad\u7ec3\u4efb\u52a1\u7684\u901a\u4fe1\u6a21\u5f0f\u3002<\/p>\n\n\n\n

\u7814\u7a76\u52a8\u673a<\/strong><\/h2>\n\n\n\n

\u56e0\u6b64\uff0c\u672c\u6587\u5bf9\u5206\u5e03\u5f0f\u8bad\u7ec3\u4efb\u52a1\u5728\u4e0d\u540c\u5e76\u884c\u7b56\u7565\u4e0b\u7684\u901a\u4fe1\u6a21\u5f0f\u8fdb\u884c\u4e86\u5206\u6790\uff0c\u5206\u522b\u662f\u6570\u636e\u5e76\u884c<\/a>\u3001\u6d41\u6c34\u7ebf\u5e76\u884c<\/a>\u3001\u5f20\u91cf\u5e76\u884c<\/a>\u4ee5\u53ca3D\u6df7\u5408\u5e76\u884c<\/a>\uff0c\u7ed3\u679c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n

\"CASSINI:<\/figure>\n\n\n\n

\u5b50\u56fe(a)-\u6570\u636e\u5e76\u884c\uff1a\u65e0\u9700\u901a\u4fe1\u7684\u524d\u53cd\u5411\u8ba1\u7b97\u4ee5\u53ca\u6700\u540e\u7684AllReduce\u901a\u4fe1\u3002<\/p>\n\n\n\n

\u5b50\u56fe(b)-\u6d41\u6c34\u7ebf\u5e76\u884c\uff1a\u4f7f\u7528PipeDream<\/a>\u6d41\u6c34\u7ebf\u5e76\u884c\u65b9\u6cd5\uff0c\u524d\u53cd\u5411\u8fc7\u7a0b\u4e2d\u6709\u6fc0\u6d3b\u503c\u7684\u901a\u4fe1\uff0c\u6700\u540e\u4f1a\u6709\u4e00\u4e2aEmbedding\u7684AllReduce\u901a\u4fe1\uff0c\u5f00\u9500\u8f83\u5927\u3002<\/p>\n\n\n\n

\u5b50\u56fe(c)-\u5f20\u91cf\u5e76\u884c\uff1a\u524d\u53cd\u5411\u8fc7\u7a0b\u90fd\u9700\u8981\u8fdb\u884c\u901a\u4fe1\uff0c\u5728\u8fed\u4ee3\u4e4b\u95f4\uff0c\u8fdb\u884c\u6570\u636e\u52a0\u8f7d\u65f6\uff0c\u6ca1\u6709\u901a\u4fe1\u5f00\u9500\u3002<\/p>\n\n\n\n

\u5b50\u56fe(d)-3D\u6df7\u5408\u5e76\u884c\uff1a\u5728\u524d\u53cd\u5411\u8fc7\u7a0b\u4e2d\u6709\u8bb8\u591a\u901a\u4fe1\u64cd\u4f5c\uff0c\u8fd9\u4e9b\u901a\u4fe1\u64cd\u4f5c\u5bf9\u7f51\u7edc\u5e26\u5bbd\u7684\u9700\u6c42\u4e0d\u540c\u3002<\/p>\n\n\n\n

\u5173\u952e\u53d1\u73b0\uff1a<\/p>\n\n\n\n