{"id":1337,"date":"2026-05-11T10:06:36","date_gmt":"2026-05-11T02:06:36","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1337"},"modified":"2026-05-11T10:14:45","modified_gmt":"2026-05-11T02:14:45","slug":"syccl-exploiting-symmetry-for-efficient-collective-communication-scheduling%e3%80%8bsigcomm-2025","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1337","title":{"rendered":"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling"},"content":{"rendered":"\n<p>SIGCOMM 2025<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>1. \u8bfe\u9898\u80cc\u666f\uff1a\u96c6\u4f53\u901a\u4fe1\u8c03\u5ea6\u7684\u74f6\u9888\u4e0e\u6311\u6218<\/strong><\/h4>\n\n\n\n<p>\u5f53\u524d\uff0c\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\uff08\u5982\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\uff09\u7684\u6548\u7387\u53d7\u5230\u96c6\u4f53\u901a\u4fe1\uff08Collective Communication\uff0c\u5982 AllReduce\u3001AllGather\uff09\u7684\u4e25\u91cd\u5236\u7ea6\u3002\u968f\u7740\u6a21\u578b\u89c4\u6a21\u4ece\u5341\u4ebf\u7ea7\u5411\u4e07\u4ebf\u7ea7\u6f14\u8fdb\uff0c\u901a\u4fe1\u8017\u65f6\u5728\u603b\u8bad\u7ec3\u65f6\u95f4\u4e2d\u7684\u5360\u6bd4\u5df2\u8d85\u8fc7 30%\u3002\u7136\u800c\uff0c\u5de5\u4e1a\u754c\u4e3b\u6d41\u5e93\uff08\u5982 NVIDIA NCCL\uff09\u4e3b\u8981\u4f9d\u8d56\u9884\u5b9a\u4e49\u7684\u3001\u56fa\u5b9a\u7684\u62d3\u6251\u7b97\u6cd5\uff08\u5982 Ring \u548c Tree\uff09\u3002\u8fd9\u4e9b\u56fa\u5b9a\u7b97\u6cd5\u5728\u9762\u5bf9\u5f02\u6784\u5e26\u5bbd\u3001\u975e\u5bf9\u79f0\u7f51\u7edc\u62d3\u6251\u6216\u7279\u5b9a\u7684\u6a21\u578b\u7b97\u5b50\u9700\u6c42\u65f6\uff0c\u5f80\u5f80\u65e0\u6cd5\u69a8\u5e72\u786c\u4ef6\u7684\u5cf0\u503c\u6027\u80fd\uff0c\u5bfc\u81f4\u6602\u8d35\u7684 H800 \u6216 A100 \u96c6\u7fa4\u51fa\u73b0\u663e\u8457\u7684\u7b97\u529b\u6d6a\u8d39\u3002<\/p>\n\n\n\n<p>\u73b0\u6709\u7684\u81ea\u52a8\u5316\u8c03\u5ea6\u5408\u6210\u5668\uff08\u5982 TACCL\u3001TECCL\uff09\u5c1d\u8bd5\u901a\u8fc7\u6df7\u5408\u6574\u6570\u7ebf\u6027\u89c4\u5212\uff08MILP\uff09\u6765\u5bfb\u6c42\u6700\u4f18\u89e3\u3002\u4f46\u8fd9\u4e9b\u65b9\u6848\u9762\u4e34\u4e25\u5cfb\u7684\u201c\u7ef4\u5ea6\u707e\u96be\u201d\uff1a\u968f\u7740 GPU \u8282\u70b9\u6570\u91cf\u7684\u589e\u52a0\uff0c\u641c\u7d22\u7a7a\u95f4\u5448\u6307\u6570\u7ea7\u7206\u70b8\u3002\u5bf9\u4e8e\u4e00\u4e2a\u62e5\u6709 128 \u4e2a GPU \u7684\u96c6\u7fa4\uff0c\u4f20\u7edf\u5408\u6210\u5668\u5f80\u5f80\u9700\u8981\u8017\u8d39\u6570\u5929\u751a\u81f3\u6570\u5468\u624d\u80fd\u7b97\u51fa\u4e00\u4e2a\u6b21\u4f18\u89e3\uff0c\u8fd9\u5728\u77ac\u606f\u4e07\u53d8\u7684\u751f\u4ea7\u73af\u5883\u4e2d\u662f\u65e0\u6cd5\u63a5\u53d7\u7684\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u5728\u5206\u949f\u7ea7\u65f6\u95f4\u5185\u4e3a\u5927\u89c4\u6a21\u96c6\u7fa4\u5408\u6210\u9ad8\u6548\u8c03\u5ea6\u65b9\u6848\uff0c\u6210\u4e3a\u4e86\u5b66\u672f\u754c\u548c\u5de5\u4e1a\u754c\u5171\u540c\u9762\u4e34\u7684\u6df1\u5c42\u79d1\u5b66\u95ee\u9898\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"597\" height=\"225\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/1.png\"  class=\"wp-image-1338\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/1.png 597w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/1-300x113.png 300w\" sizes=\"auto, (max-width: 597px) 100vw, 597px\" title=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe\" alt=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe\" \/><\/figure>\n<\/div>\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"594\" height=\"230\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/2.png\"  class=\"wp-image-1339\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/2.png 594w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/2-300x116.png 300w\" sizes=\"auto, (max-width: 594px) 100vw, 594px\" title=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe1\" alt=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe1\" \/><\/figure>\n<\/div>\n\n\n<h4 class=\"wp-block-heading\"><strong>2. \u6838\u5fc3\u6280\u672f\uff1a\u57fa\u4e8e\u5bf9\u79f0\u6027\uff08Symmetry\uff09\u7684\u89e3\u6784\u4e0e\u5408\u6210<\/strong><\/h4>\n\n\n\n<p>SYCCL \u7684\u6838\u5fc3\u8d21\u732e\u5728\u4e8e\u5176\u5bf9\u201c\u5bf9\u79f0\u6027\u201d\u7684\u6df1\u523b\u6d1e\u5bdf\u3002\u7814\u7a76\u56e2\u961f\u53d1\u73b0\uff0c\u73b0\u4ee3\u6570\u636e\u4e2d\u5fc3\u7f51\u7edc\uff08\u5982 Clos \u6216 Multi-rail \u62d3\u6251\uff09\u4ee5\u53ca\u96c6\u4f53\u901a\u4fe1\u7b97\u5b50\uff08\u5982 AllGather\uff09\u672c\u8eab\u5177\u5907\u6781\u5f3a\u7684\u7ed3\u6784\u5bf9\u79f0\u6027\u3002SYCCL \u5e76\u6ca1\u6709\u8bd5\u56fe\u53bb\u89e3\u51b3\u4e00\u4e2a\u5e9e\u5927\u7684\u5168\u5c40\u4f18\u5316\u95ee\u9898\uff0c\u800c\u662f\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u201c\u8349\u56fe\uff08Sketch\uff09\u201d\u7684\u5206\u89e3\u673a\u5236\u3002\u5b83\u5c06\u539f\u59cb\u7684\u5927\u89c4\u6a21\u901a\u4fe1\u9700\u6c42\u62c6\u5206\u4e3a\u591a\u4e2a\u89c4\u6a21\u8f83\u5c0f\u3001\u5728\u62d3\u6251\u4e0a\u5177\u6709\u5bf9\u79f0\u5173\u7cfb\u7684\u5b50\u9700\u6c42\u3002\u901a\u8fc7\u53ea\u5bf9\u4e00\u4e2a\u4ee3\u8868\u6027\u7684\u62d3\u6251\u5b50\u96c6\u8fdb\u884c\u6c42\u89e3\uff0c\u5e76\u5c06\u6240\u5f97\u65b9\u6848\u955c\u50cf\u6620\u5c04\u5230\u5176\u4ed6\u5bf9\u79f0\u90e8\u5206\uff0cSYCCL \u6781\u5927\u5730\u7f29\u51cf\u4e86\u6c42\u89e3\u96be\u5ea6\u3002<\/p>\n\n\n\n<p>\u5728\u5177\u4f53\u7684\u7b97\u6cd5\u5b9e\u73b0\u4e0a\uff0cSYCCL \u5f15\u5165\u4e86\u4e24\u9636\u6bb5\u641c\u7d22\u7b56\u7565\u3002\u7b2c\u4e00\u9636\u6bb5\u901a\u8fc7\u7c97\u7c92\u5ea6\u7684\u65f6\u95f4\u7247\u5212\u5206\u8fdb\u884c\u201c\u8349\u56fe\u201d\u7b5b\u9009\uff0c\u5feb\u901f\u5b9a\u4f4d\u5177\u6709\u9ad8\u6027\u80fd\u6f5c\u529b\u7684\u901a\u4fe1\u8def\u5f84\u7ec4\u5408\u3002\u7b2c\u4e8c\u9636\u6bb5\u5219\u9488\u5bf9\u9009\u5b9a\u7684\u8349\u56fe\u8fdb\u884c\u7ec6\u7c92\u5ea6\u7684 MILP \u4f18\u5316\uff0c\u7cbe\u786e\u6821\u51c6\u6bcf\u4e2a\u6570\u636e\u5757\u7684\u4f20\u8f93\u65f6\u523b\u548c\u8def\u5f84\u3002\u6b64\u5916\uff0cSYCCL \u8fd8\u914d\u5907\u4e86\u9ad8\u6548\u7684\u540c\u6784\u68c0\u6d4b\u5f15\u64ce\uff0c\u80fd\u591f\u81ea\u52a8\u8bc6\u522b\u62d3\u6251\u4e2d\u5197\u4f59\u7684\u8ba1\u7b97\u8def\u5f84\uff0c\u786e\u4fdd\u6c42\u89e3\u5668\u4ec5\u5728\u975e\u5197\u4f59\u7684\u7a7a\u95f4\u5185\u5de5\u4f5c\u3002\u8fd9\u79cd\u201c\u5206\u800c\u6cbb\u4e4b\u201d\u7684\u601d\u60f3\uff0c\u4f7f\u5f97 SYCCL \u80fd\u591f\u5c06\u8c03\u5ea6\u5408\u6210\u7684\u65f6\u95f4\u4ece\u6570\u5c0f\u65f6\u7f29\u77ed\u81f3\u51e0\u5206\u949f\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"570\" height=\"234\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/3.png\"  class=\"wp-image-1340\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/3.png 570w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/3-300x123.png 300w\" sizes=\"auto, (max-width: 570px) 100vw, 570px\" title=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe2\" alt=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe2\" \/><\/figure>\n<\/div>\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"554\" height=\"333\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/4.png\"  class=\"wp-image-1341\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/4.png 554w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/4-300x180.png 300w\" sizes=\"auto, (max-width: 554px) 100vw, 554px\" title=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe3\" alt=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe3\" \/><\/figure>\n<\/div>\n\n\n<h4 class=\"wp-block-heading\"><strong>3. \u5b9e\u9a8c\u9a8c\u8bc1\uff1a\u6027\u80fd\u63d0\u5347\u4e0e\u6269\u5c55\u6027\u5206\u6790<\/strong><\/h4>\n\n\n\n<p>\u5728 32 \u53f0 A100 \u670d\u52a1\u5668\u7ec4\u6210\u7684\u6d4b\u8bd5\u5e8a\u4ee5\u53ca\u57fa\u4e8e\u751f\u4ea7\u89c4\u6a21\u7684\u4eff\u771f\u5b9e\u9a8c\u4e2d\uff0cSYCCL \u5c55\u793a\u4e86\u538b\u5012\u6027\u7684\u6027\u80fd\u4f18\u52bf\u3002\u76f8\u6bd4\u4e8e\u4e1a\u754c\u6807\u6746 NCCL\uff0cSYCCL \u5728\u7279\u5b9a\u7684\u96c6\u4f53\u901a\u4fe1\u4efb\u52a1\u4e2d\u5b9e\u73b0\u4e86\u9ad8\u8fbe 127% \u7684\u541e\u5410\u91cf\u63d0\u5347\u3002\u5728\u7aef\u5230\u7aef\u7684\u5927\u6a21\u578b\u8bad\u7ec3\u6d4b\u8bd5\u4e2d\uff08\u5982 GPT-6.7B \u6a21\u578b\uff09\uff0cSYCCL \u901a\u8fc7\u4f18\u5316\u5e95\u5c42\u7684\u901a\u4fe1\u8c03\u5ea6\uff0c\u76f4\u63a5\u4e3a\u6574\u4f53\u8bad\u7ec3\u901f\u5ea6\u5e26\u6765\u4e86\u7ea6 6.3% \u7684\u63d0\u5347\u3002\u8fd9\u4e0d\u4ec5\u610f\u5473\u7740\u8bad\u7ec3\u65f6\u95f4\u7684\u7f29\u77ed\uff0c\u66f4\u610f\u5473\u7740\u5728\u540c\u7b49\u7b97\u529b\u6295\u5165\u4e0b\u80fd\u652f\u6491\u66f4\u5927\u89c4\u6a21\u7684\u4e1a\u52a1\u8fed\u4ee3\u3002<\/p>\n\n\n\n<p>\u66f4\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u662f\u5176\u5408\u6210\u6548\u7387\u7684\u8d28\u53d8\u3002\u5728\u5904\u7406 128 \u4e2a GPU \u89c4\u6a21\u7684 AllGather \u4efb\u52a1\u65f6\uff0c\u4f20\u7edf\u5408\u6210\u5668 TECCL \u56e0\u5185\u5b58\u6ea2\u51fa\u6216\u8d85\u65f6\u800c\u5931\u8d25\uff0c\u800c SYCCL \u4ec5\u9700\u4e0d\u5230 10 \u5206\u949f\u5373\u53ef\u5b8c\u6210\u9ad8\u8d28\u91cf\u65b9\u6848\u7684\u751f\u6210\u3002\u5728 H800 \u8fd9\u4e00\u7c7b\u5bf9\u5e26\u5bbd\u6781\u5176\u654f\u611f\u7684\u62d3\u6251\u7ed3\u6784\u4e2d\uff0cSYCCL \u80fd\u591f\u901a\u8fc7\u624b\u52a8\u548c\u81ea\u52a8\u7ed3\u5408\u7684\u8349\u56fe\u8bbe\u8ba1\uff0c\u5b8c\u7f8e\u9002\u914d\u5176\u7279\u6709\u7684\u5e26\u5bbd\u6bd4\u4f8b\uff0c\u6709\u6548\u89e3\u51b3\u4e86\u7531\u4e8e NVLink \u548c PCIe \u5e26\u5bbd\u4e0d\u5e73\u8861\u5e26\u6765\u7684\u901a\u4fe1\u74f6\u9888\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"561\" height=\"198\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/6.png\"  class=\"wp-image-1342\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/6.png 561w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/6-300x106.png 300w\" sizes=\"auto, (max-width: 561px) 100vw, 561px\" title=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe4\" alt=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe4\" \/><\/figure>\n<\/div>\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"570\" height=\"555\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/7.png\"  class=\"wp-image-1343\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/7.png 570w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/7-300x292.png 300w\" sizes=\"auto, (max-width: 570px) 100vw, 570px\" title=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe5\" alt=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe5\" \/><\/figure>\n<\/div>\n\n\n<h4 class=\"wp-block-heading\"><strong>4. \u7ed3\u8bba\u4e0e\u672a\u6765\u6f14\u8fdb\u5efa\u8bae<\/strong><\/h4>\n\n\n\n<p>SYCCL \u4e3a\u5927\u89c4\u6a21\u5e76\u884c\u8ba1\u7b97\u4e2d\u7684\u901a\u4fe1\u4f18\u5316\u5f00\u8f9f\u4e86\u65b0\u7684\u8303\u5f0f\u3002\u5b83\u8bc1\u660e\u4e86\u5728\u6781\u590d\u6742\u7684\u7ec4\u5408\u4f18\u5316\u95ee\u9898\u4e2d\uff0c\u5229\u7528\u5148\u9a8c\u7684\u7ed3\u6784\u5bf9\u79f0\u6027\u4fe1\u606f\u53ef\u4ee5\u4ea7\u751f\u6570\u91cf\u7ea7\u7684\u6548\u7387\u63d0\u5347\u3002SYCCL \u7684\u8f93\u51fa\u7ed3\u679c\u91c7\u7528\u4e86 XML \u683c\u5f0f\uff0c\u80fd\u591f\u4e0e\u73b0\u6709\u7684 MSCCL-executor \u7b49\u6267\u884c\u540e\u7aef\u65e0\u7f1d\u96c6\u6210\uff0c\u8fd9\u610f\u5473\u7740\u8be5\u7cfb\u7edf\u5177\u5907\u6781\u5f3a\u7684\u5de5\u4e1a\u843d\u5730\u6f5c\u529b\uff0c\u53ef\u4ee5\u4f5c\u4e3a\u5927\u6a21\u578b\u8bad\u7ec3\u5e95\u5ea7\u4e2d\u201c\u901a\u4fe1\u5f15\u64ce\u201d\u7684\u6838\u5fc3\u7ec4\u4ef6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"533\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-1024x533.png\"  class=\"wp-image-1344\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-1024x533.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-300x156.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-768x400.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9.png 1113w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe6\" alt=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe6\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"736\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-1024x736.png\"  class=\"wp-image-1345\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-1024x736.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-300x216.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-768x552.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10.png 1131w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe7\" alt=\"SYCCL: Exploiting Symmetry for Efficient Collective Communication Scheduling\u63d2\u56fe7\" \/><\/figure>\n\n\n\n<p>\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\u53ef\u4ee5\u96c6\u4e2d\u5728\u4ee5\u4e0b\u4e24\u70b9\uff1a\u4e00\u662f\u5c06 SYCCL \u6269\u5c55\u5230\u66f4\u52a0\u590d\u6742\u7684\u975e\u5bf9\u79f0\u6545\u969c\u573a\u666f\uff08\u5982\u90e8\u5206\u94fe\u8def\u635f\u574f\u65f6\u7684\u52a8\u6001\u8c03\u5ea6\uff09\uff0c\u589e\u5f3a\u7cfb\u7edf\u7684\u5bb9\u9519\u6027\u548c\u9c81\u68d2\u6027\uff1b\u4e8c\u662f\u63a2\u7d22\u5c06\u4eba\u5de5\u667a\u80fd\u6280\u672f\uff08\u5982\u5f3a\u5316\u5b66\u4e60\uff09\u4e0e SYCCL \u7684\u8349\u56fe\u641c\u7d22\u673a\u5236\u76f8\u7ed3\u5408\uff0c\u901a\u8fc7\u5b66\u4e60\u5386\u53f2\u6700\u4f18\u8c03\u5ea6\u7ecf\u9a8c\u6765\u8fdb\u4e00\u6b65\u52a0\u901f\u641c\u7d22\u8fc7\u7a0b\u3002\u6b64\u5916\uff0c\u968f\u7740 CXL \u7b49\u65b0\u578b\u4e92\u8fde\u6280\u672f\u7684\u51fa\u73b0\uff0c\u5982\u4f55\u5b9a\u4e49\u65b0\u7684\u5bf9\u79f0\u6027\u539f\u8bed\u4ee5\u652f\u6301\u5185\u5b58\u6c60\u5316\u67b6\u6784\u4e0b\u7684\u901a\u4fe1\u8c03\u5ea6\uff0c\u4e5f\u5c06\u662f\u4e00\u4e2a\u503c\u5f97\u6df1\u5165\u63a2\u8ba8\u7684\u8bfe\u9898\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>SIGCOMM 2025 1. \u8bfe\u9898\u80cc\u666f\uff1a\u96c6\u4f53\u901a\u4fe1\u8c03\u5ea6\u7684\u74f6\u9888\u4e0e\u6311\u6218 \u5f53\u524d\uff0c\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\uff08\u5982\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\uff09\u7684\u6548\u7387\u53d7\u5230\u96c6\u4f53\u901a\u4fe1\uff08Collective Communication\uff0c\u5982 AllReduce\u3001AllGather\uff09\u7684\u4e25\u91cd\u5236\u7ea6\u3002\u968f\u7740\u6a21\u578b\u89c4\u6a21\u4ece\u5341\u4ebf\u7ea7\u5411\u4e07\u4ebf\u7ea7\u6f14\u8fdb\uff0c\u901a\u4fe1\u8017\u65f6\u5728\u603b\u8bad\u7ec3\u65f6\u95f4\u4e2d\u7684\u5360\u6bd4\u5df2\u8d85\u8fc7 30%\u3002\u7136\u800c\uff0c\u5de5\u4e1a\u754c\u4e3b\u6d41\u5e93\uff08\u5982 NVIDIA NCCL\uff09\u4e3b\u8981\u4f9d\u8d56\u9884\u5b9a\u4e49\u7684\u3001\u56fa\u5b9a\u7684\u62d3\u6251\u7b97\u6cd5\uff08\u5982 Ring \u548c Tree\uff09\u3002\u8fd9\u4e9b\u56fa\u5b9a\u7b97\u6cd5\u5728\u9762\u5bf9\u5f02\u6784\u5e26\u5bbd\u3001\u975e\u5bf9\u79f0\u7f51\u7edc\u62d3\u6251\u6216\u7279\u5b9a\u7684\u6a21\u578b\u7b97 &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1337\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":5,"featured_media":1347,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[17,23,6],"tags":[13],"class_list":["post-1337","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-17","category-23","category-weilaiwangluo","tag-13"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1337","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1337"}],"version-history":[{"count":2,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1337\/revisions"}],"predecessor-version":[{"id":1362,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1337\/revisions\/1362"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1347"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1337"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1337"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1337"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}