{"id":1181,"date":"2026-03-29T14:21:01","date_gmt":"2026-03-29T06:21:01","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1181"},"modified":"2026-03-29T14:21:02","modified_gmt":"2026-03-29T06:21:02","slug":"alibaba-hpn-a-data-center-network-for-large-language-model-training","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1181","title":{"rendered":"Alibaba HPN: A Data Center Network for Large Language Model Training"},"content":{"rendered":"\n<p>SIGCOMM &#8217;24: ACM SIGCOMM 2024 Conference, August 4\u20138, 2024, Sydney, NSW, Australia<\/p>\n\n\n\n<p><a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3651890.3672265\">https:\/\/dl.acm.org\/doi\/10.1145\/3651890.3672265<\/a><\/p>\n\n\n\n<p>\u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a\uff1a\u4f20\u7edf\u6570\u636e\u4e2d\u5fc3\u7f51\u7edc\u4e3a\u4f55\u4e0d\u9002\u5408 LLM \u8bad\u7ec3<\/p>\n\n\n\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bad\u7ec3\u5bf9\u6570\u636e\u4e2d\u5fc3\u7f51\u7edc\u63d0\u51fa\u4e86\u5168\u65b0\u7684\u6311\u6218\uff0c\u4f20\u7edf\u9762\u5411\u901a\u7528\u4e91\u8ba1\u7b97\u8bbe\u8ba1\u7684\u7f51\u7edc\u67b6\u6784\u5df2\u4e0d\u518d\u9002\u7528\u3002\u963f\u91cc\u5df4\u5df4\u5728\u5b9e\u9645\u751f\u4ea7\u4e2d\u89c2\u5bdf\u5230\u4e24\u4e2a\u6839\u672c\u6027\u7684\u4e0d\u5339\u914d\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e00\u4e2a\u4e0d\u5339\u914d\u4f53\u73b0\u5728\u6d41\u91cf\u6a21\u5f0f\u4e0a\u3002\u901a\u7528\u4e91\u8ba1\u7b97\u4ea7\u751f\u6570\u767e\u4e07\u6761\u8fde\u7eed\u3001\u4f4e\u5229\u7528\u7387\u7684\u5c0f\u6d41\uff08\u901a\u5e38\u4f4e\u4e8e\u7f51\u5361\u5bb9\u91cf\u7684 20%\uff09\uff0c\u6574\u4f53\u6d41\u91cf\u6a21\u5f0f\u5e73\u7a33\u4e14\u7f13\u6162\u53d8\u5316\u3002\u800c LLM \u8bad\u7ec3\u622a\u7136\u4e0d\u540c\u2014\u2014\u6bcf\u53f0\u4e3b\u673a\u4ec5\u4ea7\u751f\u5c11\u91cf\u8fde\u63a5\uff08\u51e0\u5341\u5230\u51e0\u767e\u6761\uff09\uff0c\u4f46\u8fd9\u4e9b\u6d41\u662f\u5468\u671f\u6027\u7684\u3001\u7a81\u53d1\u6027\u7684\u5927\u8c61\u6d41\uff0c\u77ac\u95f4\u5c31\u80fd\u5360\u6ee1\u7f51\u5361\u5168\u90e8 400Gbps \u7684\u5bb9\u91cf\u3002\u8fd9\u79cd\u4f4e\u71b5\u3001\u9ad8\u7a81\u53d1\u7684\u6d41\u91cf\u7279\u5f81\u4f7f\u5f97\u4f20\u7edf\u6570\u636e\u4e2d\u5fc3\u5e7f\u6cdb\u91c7\u7528\u7684 ECMP \u8d1f\u8f7d\u5747\u8861\u65b9\u6848\u5931\u6548\u3002ECMP \u4f9d\u8d56\u5927\u91cf\u6d41\u7684\u54c8\u5e0c\u503c\u5728\u591a\u6761\u7b49\u4ef7\u8def\u5f84\u4e0a\u5747\u5300\u5206\u5e03\uff0c\u800c\u5f53\u6d41\u7684\u6570\u91cf\u6781\u5c11\u4e14\u6bcf\u6761\u6d41\u90fd\u662f\u5927\u8c61\u6d41\u65f6\uff0c\u54c8\u5e0c\u6781\u5316\uff08hash polarization\uff09\u5bfc\u81f4\u67d0\u4e9b\u94fe\u8def\u4e25\u91cd\u62e5\u585e\u800c\u5176\u4ed6\u94fe\u8def\u7a7a\u95f2\u3002\u66f4\u4e25\u91cd\u7684\u662f\uff0c\u4f20\u7edf\u4e09\u5c42 Clos \u67b6\u6784\u4e2d\u5927\u8c61\u6d41\u9700\u8981\u7ecf\u8fc7\u4e09\u6b21\u7ea7\u8054\u54c8\u5e0c\uff08ToR\u3001\u6c47\u805a\u5c42\u3001\u6838\u5fc3\u5c42\uff09\uff0c\u52a0\u5267\u4e86\u8d1f\u8f7d\u4e0d\u5747\u8861\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e8c\u4e2a\u4e0d\u5339\u914d\u4f53\u73b0\u5728\u6545\u969c\u654f\u611f\u6027\u4e0a\u3002LLM \u8bad\u7ec3\u662f\u4e00\u4e2a\u540c\u6b65\u8fc7\u7a0b\uff0c\u6240\u6709 GPU \u534f\u540c\u5b8c\u6210\u6bcf\u4e2a\u8fed\u4ee3\uff0c\u4efb\u4f55\u4e00\u5757 GPU \u6216\u4e3b\u673a\u7684\u5f02\u5e38\u90fd\u53ef\u80fd\u5ef6\u8fdf\u751a\u81f3\u5d29\u6e83\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u3002\u751f\u4ea7\u6570\u636e\u663e\u793a\uff0cLLM \u8bad\u7ec3\u4e2d\u7684\u6545\u969c\u6210\u672c\u662f\u901a\u7528\u4e91\u8ba1\u7b97\u7684 20 \u500d\u2014\u2014\u4e00\u4e2a\u4f7f\u7528 3000 \u5757 GPU \u7684\u8bad\u7ec3\u4efb\u52a1\u6bcf\u5c0f\u65f6\u6210\u672c\u7ea6 2 \u4e07\u7f8e\u5143\uff0c\u4e00\u6b21\u6545\u969c\u56de\u9000\u5230\u6570\u5c0f\u65f6\u524d\u7684\u68c0\u67e5\u70b9\u610f\u5473\u7740\u7ea6 3 \u4e07\u7f8e\u5143\u7684\u76f4\u63a5\u635f\u5931\u3002\u800c ToR \u4ea4\u6362\u673a\u7684\u5355\u70b9\u6545\u969c\u662f\u6700\u81f4\u547d\u7684\uff0c\u56e0\u4e3a\u4e00\u53f0 ToR \u6545\u969c\u4f1a\u5bfc\u81f4\u6570\u5341\u751a\u81f3\u4e0a\u767e\u53f0\u4e3b\u673a\u4e0d\u53ef\u7528\u3002\u963f\u91cc\u5df4\u5df4\u7684\u8fd0\u8425\u6570\u636e\u663e\u793a\uff0c\u6bcf\u6708\u7ea6 0.057% \u7684 NIC-ToR \u94fe\u8def\u6545\u969c\uff0c\u7ea6 0.051% \u7684 ToR \u4ea4\u6362\u673a\u53d1\u751f\u4e25\u91cd\u9519\u8bef\uff0c\u6bcf\u5929\u8fd8\u6709 5000-60000 \u6b21\u94fe\u8def\u6296\u52a8\u4e8b\u4ef6\u3002<\/p>\n\n\n\n<p>\u4e8c\u3001HPN \u6574\u4f53\u67b6\u6784\uff1a\u4e24\u5c42\u53cc\u5e73\u9762\u53d6\u4ee3\u4e09\u5c42 Clos<\/p>\n\n\n\n<p>\u57fa\u4e8e\u4e0a\u8ff0\u6311\u6218\uff0c\u963f\u91cc\u5df4\u5df4\u8bbe\u8ba1\u5e76\u90e8\u7f72\u4e86 HPN\uff08High Performance Network\uff09\uff0c\u4e00\u79cd\u4e13\u4e3a LLM \u8bad\u7ec3\u6253\u9020\u7684\u6570\u636e\u4e2d\u5fc3\u7f51\u7edc\u67b6\u6784\u3002HPN \u7684\u76ee\u6807\u6709\u4e09\u4e2a\uff1a\u6269\u5c55\u6027\uff08\u5355 Pod \u652f\u6301 15000 \u5757 GPU\uff09\u3001\u9ad8\u6027\u80fd\uff08\u6700\u5c0f\u5316\u7f51\u7edc\u8df3\u6570\u548c ECMP \u54c8\u5e0c\u6b21\u6570\uff09\u3001\u4ee5\u53ca ToR \u5355\u70b9\u6545\u969c\u5bb9\u9519\u3002<\/p>\n\n\n\n<p>HPN \u5c06\u7f51\u7edc\u5206\u4e3a\u524d\u7aef\u7f51\u7edc\u548c\u540e\u7aef\u7f51\u7edc\u3002\u540e\u7aef\u7f51\u7edc\u627f\u8f7d\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u901a\u4fe1\u6d41\u91cf\uff0c\u524d\u7aef\u7f51\u7edc\u5904\u7406\u7ba1\u7406\u3001\u63a8\u7406\u548c\u5b58\u50a8\u7b49\u5176\u4ed6\u6d41\u91cf\uff0c\u4e24\u8005\u7269\u7406\u9694\u79bb\u4ee5\u786e\u4fdd\u8bad\u7ec3\u6d41\u91cf\u4e0d\u53d7\u5e72\u6270\u3002\u6bcf\u53f0\u4e3b\u673a\u914d\u5907 8 \u5757 GPU\uff0c\u901a\u8fc7 NVLink \u9ad8\u5e26\u5bbd\u4e92\u8fde\uff08400-900GBps \u53cc\u5411\uff09\uff0c\u540c\u65f6\u914d\u5907 9 \u5757\u53cc\u7aef\u53e3 200Gbps \u7f51\u5361\uff0c\u5176\u4e2d 8 \u5757\u5206\u522b\u670d\u52a1 8 \u5757 GPU \u63a5\u5165\u540e\u7aef\u7f51\u7edc\uff08\u6bcf\u5757 GPU \u72ec\u4eab 400Gbps RDMA \u5e26\u5bbd\uff09\uff0c1 \u5757\u63a5\u5165\u524d\u7aef\u7f51\u7edc\u3002\u6bcf\u5757\u7f51\u5361\u7684\u4e24\u4e2a\u7aef\u53e3\u5206\u522b\u8fde\u63a5\u5230\u4e0d\u540c\u7684 ToR \u4ea4\u6362\u673a\uff0c\u5f62\u6210\u53cc ToR \u8bbe\u8ba1\u3002<\/p>\n\n\n\n<p>HPN \u7684\u540e\u7aef\u7f51\u7edc\u91c7\u7528\u4e24\u5c42\u53cc\u5e73\u9762\u67b6\u6784\u800c\u975e\u4f20\u7edf\u7684\u4e09\u5c42 Clos\uff0c\u901a\u8fc7\u56db\u4e2a\u5173\u952e\u673a\u5236\u9010\u6b65\u6269\u5927\u7f51\u7edc\u89c4\u6a21\uff1a\u53cc ToR \u4f7f\u89c4\u6a21\u7ffb\u500d\uff0c51.2Tbps \u5355\u82af\u7247\u4ea4\u6362\u673a\u5145\u5206\u5229\u7528\u7aef\u53e3\u5bb9\u91cf\uff0cRail-Optimized \u62d3\u6251\u4f7f\u89c4\u6a21\u6269\u5927 8 \u500d\uff0c\u53cc\u5e73\u9762\u8bbe\u8ba1\u518d\u6b21\u4f7f\u89c4\u6a21\u7ffb\u500d\u3002\u6700\u7ec8\u5b9e\u73b0\u5355\u5c42\u7f51\u7edc\uff08Segment\uff09\u5bb9\u7eb3 1024 \u5757 GPU\uff0c\u5355 Pod \u5bb9\u7eb3 15000 \u5757 GPU\u2014\u2014\u8fd9\u4e00\u89c4\u6a21\u5728\u4f20\u7edf\u67b6\u6784\u4e2d\u9700\u8981\u4e09\u5c42 Clos \u624d\u80fd\u5b9e\u73b0\u3002\u751f\u4ea7\u6570\u636e\u663e\u793a 96.3% \u7684\u8bad\u7ec3\u4efb\u52a1\u4f7f\u7528\u4e0d\u8d85\u8fc7 1000 \u5757 GPU\uff0c\u56e0\u6b64\u7edd\u5927\u591a\u6570\u4efb\u52a1\u53ef\u4ee5\u5728\u5355\u4e2a Segment \u5185\u5b8c\u6210\uff0c\u4eab\u53d7\u6700\u4f18\u7f51\u7edc\u6027\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"867\" height=\"313\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-36.png\"  class=\"wp-image-1182\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-36.png 867w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-36-300x108.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-36-768x277.png 768w\" sizes=\"auto, (max-width: 867px) 100vw, 867px\" title=\"Alibaba HPN: A Data Center Network for Large Language Model Training\u63d2\u56fe\" alt=\"Alibaba HPN: A Data Center Network for Large Language Model Training\u63d2\u56fe\" \/><\/figure>\n\n\n\n<p>\u4e09\u3001\u975e\u5806\u53e0\u53cc ToR\uff1a\u4ece\u6839\u6e90\u6d88\u9664\u5355\u70b9\u6545\u969c<\/p>\n\n\n\n<p>\u4f20\u7edf\u6570\u636e\u4e2d\u5fc3\u666e\u904d\u91c7\u7528\u5355 ToR \u8bbe\u8ba1\uff0c\u5373\u6bcf\u5757\u7f51\u5361\u7684\u4e24\u4e2a\u7aef\u53e3\u901a\u8fc7\u4e00\u6839\u7ebf\u7f06\u8fde\u63a5\u5230\u540c\u4e00\u53f0 ToR \u4ea4\u6362\u673a\u3002\u8fd9\u79cd\u8bbe\u8ba1\u5728 ToR \u6545\u969c\u65f6\u4f1a\u5bfc\u81f4\u5176\u4e0b\u6240\u6709\u4e3b\u673a\u65ad\u8fde\u3002\u4e1a\u754c\u5df2\u6709\u7684\u5806\u53e0\u53cc ToR \u65b9\u6848\uff08\u5982 vPC\u3001M-LAG\uff09\u5c06\u4e24\u53f0 ToR \u901a\u8fc7\u76f4\u8fde\u94fe\u8def\u540c\u6b65\u72b6\u6001\uff0c\u770b\u4f3c\u80fd\u89e3\u51b3\u95ee\u9898\uff0c\u4f46\u963f\u91cc\u5df4\u5df4\u5728\u4e09\u5e74\u751f\u4ea7\u5b9e\u8df5\u4e2d\u53d1\u73b0\uff0c\u5806\u53e0\u53cc ToR \u5f15\u5165\u7684\u6545\u969c\u53cd\u800c\u5360\u5230\u4e86\u4f20\u7edf\u6570\u636e\u4e2d\u5fc3\u5168\u90e8\u4e25\u91cd\u6545\u969c\u7684 40% \u4ee5\u4e0a\u3002\u5178\u578b\u95ee\u9898\u5305\u62ec\uff1a\u4e00\u53f0 ToR \u6570\u636e\u9762\u6545\u969c\u4f46\u63a7\u5236\u9762\u6b63\u5e38\u65f6\uff0c\u53e6\u4e00\u53f0 ToR \u56e0\u65e0\u6cd5\u540c\u6b65\u800c\u88ab\u8feb\u5173\u95ed\u81ea\u8eab\uff0c\u5bfc\u81f4\u6574\u673a\u67b6\u5b95\u673a\uff1b\u53cc ToR \u5347\u7ea7\u8fc7\u7a0b\u4e2d\u65b0\u65e7\u7248\u672c RPC \u4e0d\u517c\u5bb9\uff0c70% \u7684\u5347\u7ea7\u65e0\u6cd5\u6ee1\u8db3 ISSU \u7684\u7248\u672c\u5dee\u5f02\u5047\u8bbe\u3002<\/p>\n\n\n\n<p>HPN \u63d0\u51fa\u4e86\u975e\u5806\u53e0\u53cc ToR \u65b9\u6848\uff0c\u5f7b\u5e95\u53bb\u6389\u4e86\u4e24\u53f0 ToR \u4e4b\u95f4\u7684\u76f4\u8fde\u94fe\u8def\uff0c\u4f7f\u4e24\u53f0 ToR \u5b8c\u5168\u72ec\u7acb\u8fd0\u884c\u3002\u5b9e\u73b0\u8fd9\u4e00\u65b9\u6848\u7684\u6838\u5fc3\u96be\u9898\u5728\u4e8e\u5982\u4f55\u8ba9\u4e24\u53f0\u72ec\u7acb\u7684 ToR \u5728\u4e3b\u673a\u770b\u6765\u50cf\u662f\u4e00\u53f0\u8bbe\u5907\u3002HPN \u901a\u8fc7\u4e0e\u4ea4\u6362\u673a\u5382\u5546\u6df1\u5ea6\u5408\u4f5c\uff0c\u5b9a\u5236\u4e86 LACP \u6a21\u5757\uff1a\u4e24\u53f0 ToR \u4f7f\u7528 RFC \u9884\u7559\u7684\u865a\u62df\u8def\u7531\u5668 MAC \u5730\u5740\u4f5c\u4e3a\u7edf\u4e00\u7684\u7cfb\u7edf\u6807\u8bc6\uff0c\u5e76\u901a\u8fc7\u7aef\u53e3\u504f\u79fb\u7b97\u6cd5\u751f\u6210\u4e0d\u540c\u7684\u7aef\u53e3 ID\u3002\u5728\u6545\u969c\u5904\u7406\u65b9\u9762\uff0cHPN \u5c06\u6240\u6709 ARP \u8f6c\u6362\u4e3a \/32 \u4e3b\u673a\u8def\u7531\u6ce8\u5165 BGP\uff0c\u5f53\u67d0\u6761 NIC-ToR \u94fe\u8def\u6545\u969c\u65f6\uff0c\u5bf9\u5e94\u7684\u4e3b\u673a\u8def\u7531\u88ab\u64a4\u56de\uff0cBGP \u6536\u655b\u81ea\u52a8\u5c06\u6d41\u91cf\u5207\u6362\u5230\u53e6\u4e00\u53f0 ToR\u3002\u540c\u65f6\u5173\u95ed\u4e8c\u5c42\u5e7f\u64ad\u5e76\u5b9e\u73b0 ARP \u4ee3\u7406\uff0c\u786e\u4fdd Segment \u5185\u90e8\u6d41\u91cf\u4e5f\u80fd\u901a\u8fc7\u4e09\u5c42\u8def\u7531\u6b63\u786e\u8f6c\u53d1\u3002<\/p>\n\n\n\n<p>\u56db\u3001Tier1 \u8bbe\u8ba1\uff1a\u5355 Segment \u5bb9\u7eb3 1024 \u5757 GPU<\/p>\n\n\n\n<p>HPN \u5728 Tier1 \u5c42\u91c7\u7528 51.2Tbps \u5355\u82af\u7247\u4ea4\u6362\u673a\u3002\u9009\u62e9\u5355\u82af\u7247\u800c\u975e\u591a\u82af\u7247\u673a\u6846\u5f0f\u4ea4\u6362\u673a\u7684\u539f\u56e0\u662f\uff1a\u963f\u91cc\u5df4\u5df4\u7684\u957f\u671f\u8fd0\u8425\u7ecf\u9a8c\u8868\u660e\uff0c\u5c3d\u7ba1\u5355\u82af\u7247\u4ea4\u6362\u673a\u6570\u91cf\u662f\u591a\u82af\u7247\u4ea4\u6362\u673a\u7684 32.6 \u500d\uff0c\u4f46\u591a\u82af\u7247\u4ea4\u6362\u673a\u7684\u4e25\u91cd\u786c\u4ef6\u6545\u969c\u603b\u6570\u53cd\u800c\u9ad8\u51fa 3.77 \u500d\u3002\u6839\u56e0\u5728\u4e8e\u591a\u82af\u7247\u4ea4\u6362\u673a\u672c\u8d28\u4e0a\u662f\u5206\u5e03\u5f0f\u4ea4\u6362\u7cfb\u7edf\uff0c\u5185\u90e8 Fabric\u3001\u82af\u7247\u95f4\u4ea4\u4e92\u548c\u82af\u7247-CPU \u901a\u4fe1\u90fd\u53ef\u80fd\u5f15\u53d1\u6545\u969c\u3002<\/p>\n\n\n\n<p>51.2Tbps \u82af\u7247\u5e26\u6765\u4e86\u6563\u70ed\u6311\u6218\u2014\u2014\u529f\u8017\u6bd4\u4e0a\u4e00\u4ee3 25.6Tbps \u82af\u7247\u589e\u52a0\u4e86 45%\uff0c\u800c\u6700\u5927\u7ed3\u6e29\u4fdd\u6301\u4e0d\u53d8\u3002\u73b0\u6709\u7684\u70ed\u7ba1\u548c\u6807\u51c6\u5747\u6e29\u677f\u65b9\u6848\u90fd\u65e0\u6cd5\u652f\u6491\u82af\u7247\u5728\u5168\u529f\u7387\u4e0b\u6301\u7eed\u8fd0\u884c\u3002HPN \u8bbe\u8ba1\u4e86\u4f18\u5316\u7684\u5747\u6e29\u677f\u6563\u70ed\u5668\uff0c\u901a\u8fc7\u6539\u8fdb\u82af\u7247\u4e2d\u5fc3\u533a\u57df\u7684\u6bdb\u7ec6\u7ed3\u6784\u548c\u589e\u52a0\u70e7\u7ed3\u94dc\u67f1\uff0c\u4f7f\u6563\u70ed\u6548\u7387\u63d0\u5347 15%\uff0c\u5f7b\u5e95\u89e3\u51b3\u4e86\u8fc7\u6e29\u4fdd\u62a4\u5bfc\u81f4\u7684\u505c\u673a\u98ce\u9669\u3002<\/p>\n\n\n\n<p>\u7ed3\u5408 Rail-Optimized \u62d3\u6251\uff0c\u540c\u4e00 Rail \u7684\u7f51\u5361\u8fde\u63a5\u5230\u540c\u4e00\u7ec4\u53cc ToR \u4ea4\u6362\u673a\uff0c\u4e0d\u540c Rail \u4e4b\u95f4\u901a\u8fc7\u4e3b\u673a\u5185\u90e8 NVLink \u8fdb\u884c\u4e2d\u8f6c\u3002\u6bcf\u7ec4\u53cc ToR \u670d\u52a1 128 \u5757 GPU\uff0c16 \u53f0 ToR \u5171\u540c\u8fde\u63a5 1024 \u5757 GPU \u6784\u6210\u4e00\u4e2a Segment\uff0c\u5927\u5e45\u51cf\u5c11\u4e86\u9700\u8981\u8de8\u6c47\u805a\u5c42\u7684\u6d41\u91cf\u3002<\/p>\n\n\n\n<p>\u4e94\u3001Tier2 \u8bbe\u8ba1\uff1a\u53cc\u5e73\u9762\u6d88\u9664\u54c8\u5e0c\u6781\u5316<\/p>\n\n\n\n<p>\u5982\u679c\u5728 Tier2 \u5c42\u7b80\u5355\u90e8\u7f72\u5178\u578b Clos \u62d3\u6251\uff0c\u53cc ToR \u5e26\u6765\u7684\u6d41\u91cf\u6c47\u805a\u4ecd\u7136\u4f1a\u5bfc\u81f4\u54c8\u5e0c\u6781\u5316\u3002HPN \u7684\u751f\u4ea7\u6d4b\u91cf\u663e\u793a\uff0c\u5728\u8bad\u7ec3 GPT-3 175B \u53d8\u4f53\u65f6\uff0c\u53cc ToR \u4e2d\u4e24\u4e2a\u4e0b\u884c\u7aef\u53e3\u7684\u8d1f\u8f7d\u5dee\u5f02\u53ef\u8fbe 3 \u500d\u3002<\/p>\n\n\n\n<p>HPN \u7684\u89e3\u51b3\u65b9\u6848\u662f\u53cc\u5e73\u9762\u8bbe\u8ba1\uff1a\u5c06\u6bcf\u7ec4\u53cc ToR \u4e2d\u7684\u4e24\u53f0\u4ea4\u6362\u673a\u5206\u522b\u5f52\u5165\u4e24\u4e2a\u72ec\u7acb\u7684\u7f51\u7edc\u5e73\u9762\u3002\u4e00\u65e6\u6d41\u91cf\u4ece\u6e90\u7f51\u5361\u7684\u67d0\u4e2a\u7aef\u53e3\u8fdb\u5165\u67d0\u4e2a\u5e73\u9762\uff0c\u5176\u5728 Pod \u5185\u7684\u8f6c\u53d1\u8def\u5f84\u5c31\u5b8c\u5168\u786e\u5b9a\uff0c\u4e0d\u518d\u7ecf\u8fc7\u4efb\u4f55 ECMP \u54c8\u5e0c\u3002\u90e8\u7f72\u53cc\u5e73\u9762\u540e\uff0c\u4e24\u4e2a\u7aef\u53e3\u7684\u5165\u5411\u6d41\u91cf\u53d8\u5f97\u5747\u8861\uff0cToR \u4e0b\u884c\u7aef\u53e3\u7684\u961f\u5217\u957f\u5ea6\u4e0b\u964d\u4e86 91.8%\uff0c\u8de8 Segment \u6d41\u91cf\u6027\u80fd\u63d0\u5347\u4e86 71.6%\u3002<\/p>\n\n\n\n<p>\u53cc\u5e73\u9762\u8fd8\u5927\u5e45\u7b80\u5316\u4e86\u8def\u5f84\u9009\u62e9\u7684\u641c\u7d22\u7a7a\u95f4\u3002\u5728 HPN \u4e2d\uff0c\u641c\u7d22\u4e0d\u76f8\u4ea4\u8def\u5f84\u4ec5\u9700\u68c0\u67e5\u6bcf\u53f0 ToR \u4ea4\u6362\u673a\u7684 60 \u6761\u4e0a\u884c\u94fe\u8def\uff0c\u590d\u6742\u5ea6\u4e3a O(60)\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cNVIDIA SuperPod \u7684\u4e09\u5c42\u67b6\u6784\u9700\u8981\u641c\u7d22 O(4096) \u6761\u8def\u5f84\uff0cGoogle Jupiter \u9700\u8981\u641c\u7d22 O(2048) \u6761\u8def\u5f84\u3002HPN \u5728\u96c6\u5408\u901a\u4fe1\u5e93\u4e2d\u5b9e\u73b0\u4e86\u57fa\u4e8e\u4e0d\u76f8\u4ea4\u8def\u5f84\u7684\u8d1f\u8f7d\u5747\u8861\u65b9\u6848\uff1a\u5229\u7528 RePaC \u6280\u672f\u7cbe\u786e\u8ba1\u7b97\u6bcf\u6761\u8def\u5f84\u7684\u54c8\u5e0c\u7ed3\u679c\uff0c\u5efa\u7acb\u591a\u6761\u4e0d\u76f8\u4ea4\u7684 RDMA \u8fde\u63a5\uff0c\u5e76\u901a\u8fc7\u8ddf\u8e2a\u6bcf\u6761\u8fde\u63a5\u4e0a\u672a\u5b8c\u6210\u7684 WQE \u5b57\u8282\u6570\u6765\u9009\u62e9\u6700\u7a7a\u95f2\u7684\u8def\u5f84\u3002\u5728 512 \u5757 GPU \u4e0a\u5e76\u53d1\u8fd0\u884c\u56db\u4e2a AllReduce \u4efb\u52a1\u7684\u6d4b\u8bd5\u4e2d\uff0c\u8fd9\u4e00\u4f18\u5316\u8def\u5f84\u9009\u62e9\u65b9\u6848\u5c06\u96c6\u5408\u901a\u4fe1\u6027\u80fd\u63d0\u5347\u4e86 34.7%\u3002<\/p>\n\n\n\n<p>\u516d\u3001\u5b9e\u9a8c\u8bc4\u4f30\uff1a\u8bad\u7ec3\u541e\u5410\u91cf\u63d0\u5347 14.9%<\/p>\n\n\n\n<p>HPN \u5df2\u5728\u963f\u91cc\u4e91\u751f\u4ea7\u73af\u5883\u4e2d\u90e8\u7f72\u8d85\u8fc7\u516b\u4e2a\u6708\uff0c\u670d\u52a1\u6570\u5341\u5bb6\u5ba2\u6237\u7684\u6570\u5343\u4e2a\u6a21\u578b\u8bad\u7ec3\u4efb\u52a1\u3002\u8bba\u6587\u5c06 HPN \u4e0e\u4e0a\u4e00\u4ee3\u8bad\u7ec3\u7f51\u7edc\u67b6\u6784 DCN+\uff08\u5e26\u53cc ToR \u7684\u4f20\u7edf\u4e09\u5c42 Clos \u5168\u5bf9\u5206\u5e26\u5bbd\u7f51\u7edc\uff09\u8fdb\u884c\u4e86\u5bf9\u6bd4\u3002<\/p>\n\n\n\n<p>\u5728\u7aef\u5230\u7aef\u8bad\u7ec3\u6027\u80fd\u65b9\u9762\uff0c\u963f\u91cc\u4e91\u4e00\u4e2a\u81ea\u6709\u7684\u5927\u6a21\u578b\u5728 2300 \u591a\u5757 GPU \u4e0a\u8bad\u7ec3\uff0c\u4ece DCN+ \u8fc1\u79fb\u5230 HPN \u540e\uff0c\u7aef\u5230\u7aef\u8bad\u7ec3\u6027\u80fd\u63d0\u5347\u4e86 14.9%\u3002\u6c47\u805a\u5c42\u4ea4\u6362\u673a\u7684\u7edf\u8ba1\u663e\u793a\u8de8 Segment \u6d41\u91cf\u5e73\u5747\u51cf\u5c11\u4e86 37%\uff0c\u961f\u5217\u62e5\u585e\u663e\u8457\u7f13\u89e3\u3002\u5728\u4ee3\u8868\u6027\u6a21\u578b\u7684\u8bad\u7ec3\u4e2d\uff0cLLaMa-7B\u3001LLaMa-13B \u548c GPT3-175B \u5206\u522b\u83b7\u5f97\u4e86 7.9%\u300114.4% \u548c 6.3% \u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n\n\n\n<p>\u5728\u96c6\u5408\u901a\u4fe1\u6027\u80fd\u65b9\u9762\uff0cHPN \u5728 448 \u5757 GPU \u4e0a\u5c06 AllReduce \u6027\u80fd\u63d0\u5347\u4e86\u6700\u9ad8 59.3%\uff0cMulti-AllReduce\uff08\u7528\u4e8e TP=8 \u7684\u68af\u5ea6\u540c\u6b65\uff09\u6027\u80fd\u63d0\u5347\u4e86\u6700\u9ad8 158.2%\u3002<\/p>\n\n\n\n<p>\u5728\u53ef\u9760\u6027\u65b9\u9762\uff0c\u516b\u4e2a\u6708\u7684\u751f\u4ea7\u8fd0\u8425\u4e2d\u672a\u89c2\u5bdf\u5230\u4efb\u4f55 ToR \u76f8\u5173\u7684\u5355\u70b9\u6545\u969c\u3002\u6ce8\u5165\u94fe\u8def\u6545\u969c\u6d4b\u8bd5\u663e\u793a\uff0c\u5355 ToR \u8bbe\u8ba1\u4e0b\u94fe\u8def\u6545\u969c\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u7acb\u5373\u505c\u6b62\u4e14\u53ef\u80fd\u65e0\u6cd5\u6062\u590d\uff0c\u800c\u53cc ToR \u8bbe\u8ba1\u4e0b\u4ec5\u4ea7\u751f 6.25% \u7684\u6027\u80fd\u4e0b\u964d\uff0c\u4fee\u590d\u540e\u7acb\u5373\u6062\u590d\u6b63\u5e38\u3002\u94fe\u8def\u6296\u52a8\u6d4b\u8bd5\u4e2d\uff0c\u5355 ToR \u5bfc\u81f4\u8bad\u7ec3\u505c\u987f\u8d85\u8fc7 9 \u79d2\uff0c\u800c\u53cc ToR \u7684\u6027\u80fd\u5f71\u54cd\u51e0\u4e4e\u53ef\u4ee5\u5ffd\u7565\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"878\" height=\"188\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-37.png\"  class=\"wp-image-1183\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-37.png 878w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-37-300x64.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-37-768x164.png 768w\" sizes=\"auto, (max-width: 878px) 100vw, 878px\" title=\"Alibaba HPN: A Data Center Network for Large Language Model Training\u63d2\u56fe1\" alt=\"Alibaba HPN: A Data Center Network for Large Language Model Training\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<p>\u4e03\u3001\u8bbe\u8ba1\u53d6\u820d\u4e0e\u7ecf\u9a8c\u6559\u8bad<\/p>\n\n\n\n<p>\u8bba\u6587\u5206\u4eab\u4e86\u591a\u9879\u6709\u4ef7\u503c\u7684\u8bbe\u8ba1\u51b3\u7b56\u548c\u5b9e\u6218\u7ecf\u9a8c\u3002\u5173\u4e8e\u4e3a\u4f55\u4e0d\u5728 Tier2 \u91c7\u7528 Rail-Only \u62d3\u6251\u6765\u8fdb\u4e00\u6b65\u6269\u5927\u89c4\u6a21\uff1a\u867d\u7136 Rail-Only \u53ef\u4ee5\u5c06\u5355 Pod \u6269\u5c55\u5230 12 \u4e07\u5757 GPU\uff0c\u4f46\u5b83\u8981\u6c42\u6240\u6709\u6d41\u91cf\u90fd\u5728\u540c\u4e00 Rail \u5185\uff0c\u8fd9\u5728 MoE \u6a21\u578b\u8bad\u7ec3\uff08\u9700\u8981\u8de8 Rail \u7684 All-to-All \u901a\u4fe1\uff09\u548c\u591a\u79df\u6237\u573a\u666f\u4e0b\u662f\u4e0d\u53ef\u63a5\u53d7\u7684\u9650\u5236\u3002\u5173\u4e8e\u5b58\u50a8\u96c6\u7fa4\u7684\u4f4d\u7f6e\uff1a\u867d\u7136\u540e\u7aef\u7f51\u7edc\u5e26\u5bbd\u66f4\u9ad8\uff083.2Tbps vs 400Gbps\uff09\uff0c\u4f46\u5c06\u5b58\u50a8\u653e\u5728\u540e\u7aef\u7f51\u7edc\u4f1a\u5f15\u5165\u8de8\u7f51\u7edc\u4ee3\u7406\u7684\u590d\u6742\u6027\u3001\u5f71\u54cd\u8bad\u7ec3\u6027\u80fd\u6ce2\u52a8\u3001\u5e76\u5360\u7528\u5b9d\u8d35\u7684 ToR \u7aef\u53e3\uff0c\u56e0\u6b64\u6700\u7ec8\u9009\u62e9\u653e\u5728\u524d\u7aef\u7f51\u7edc\u3002\u5173\u4e8e\u5e03\u7ebf\u590d\u6742\u6027\uff1aHPN \u7684 Rail-Optimized \u548c\u53cc\u5e73\u9762\u8bbe\u8ba1\u4f7f\u5e03\u7ebf\u66f4\u52a0\u590d\u6742\uff0c\u521d\u671f\u65bd\u5de5\u4eba\u5458\u9891\u7e41\u63a5\u9519\u7ebf\u7f06\uff0c\u9700\u8981\u901a\u8fc7 INT \u63a2\u6d4b\u9010\u8df3\u9a8c\u8bc1\u6bcf\u6761\u8def\u5f84\u662f\u5426\u4e0e\u84dd\u56fe\u4e00\u81f4\u3002<\/p>\n\n\n\n<p>\u4e00\u4e2a\u503c\u5f97\u5173\u6ce8\u7684\u7269\u7406\u8bbe\u8ba1\u51b3\u7b56\u662f\u5c06\u5355\u4e2a Pod \u5b8c\u6574\u5bb9\u7eb3\u5728\u4e00\u680b\u6570\u636e\u4e2d\u5fc3\u5efa\u7b51\u5185\u3002\u963f\u91cc\u4e91\u6bcf\u680b\u5efa\u7b51\u7684\u529f\u7387\u7ea6\u675f\u4e3a 18MW\uff0c\u6070\u597d\u53ef\u4ee5\u5bb9\u7eb3\u7ea6 15000 \u5757 GPU\u3002\u8fd9\u4f7f\u5f97 Pod \u5185\u6240\u6709\u5149\u7ea4\u957f\u5ea6\u63a7\u5236\u5728 100 \u7c73\u4ee5\u5185\uff0c\u53ef\u4ee5\u4f7f\u7528\u6210\u672c\u4f4e 70% \u7684\u591a\u6a21\u5149\u6536\u53d1\u5668\u66ff\u4ee3\u5355\u6a21\u5149\u6536\u53d1\u5668\u3002<\/p>\n\n\n\n<p>\u516b\u3001\u603b\u7ed3\u4e0e\u610f\u4e49<\/p>\n\n\n\n<p>HPN \u662f\u76ee\u524d\u516c\u5f00\u53d1\u8868\u7684\u6700\u5b8c\u6574\u7684 LLM \u8bad\u7ec3\u4e13\u7528\u6570\u636e\u4e2d\u5fc3\u7f51\u7edc\u67b6\u6784\u4e4b\u4e00\u3002\u5b83\u7684\u6838\u5fc3\u6d1e\u5bdf\u662f\uff1aLLM \u8bad\u7ec3\u7684\u6d41\u91cf\u6a21\u5f0f\u4e0e\u901a\u7528\u4e91\u8ba1\u7b97\u5b58\u5728\u6839\u672c\u6027\u5dee\u5f02\uff0c\u7f51\u7edc\u67b6\u6784\u5fc5\u987b\u9488\u5bf9\u4f4e\u71b5\u3001\u9ad8\u7a81\u53d1\u3001\u5c11\u8fde\u63a5\u7684\u5927\u8c61\u6d41\u7279\u5f81\u8fdb\u884c\u4e13\u95e8\u8bbe\u8ba1\u3002\u901a\u8fc7\u975e\u5806\u53e0\u53cc ToR\u3001Rail-Optimized \u62d3\u6251\u548c\u53cc\u5e73\u9762\u67b6\u6784\u7684\u534f\u540c\u8bbe\u8ba1\uff0cHPN \u5728\u964d\u4f4e\u7f51\u7edc\u6210\u672c\u7ea6 30% \u7684\u540c\u65f6\u5c06\u8bad\u7ec3\u541e\u5410\u91cf\u63d0\u5347\u4e86 14.9%\uff0c\u5e76\u4ece\u6839\u6e90\u4e0a\u6d88\u9664\u4e86 ToR \u5355\u70b9\u6545\u969c\u3002\u5bf9\u4e8e\u6784\u5efa\u4e0b\u4e00\u4ee3 AI \u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\uff0cHPN \u5728\u62d3\u6251\u8bbe\u8ba1\u3001\u6545\u969c\u5bb9\u9519\u548c\u8d1f\u8f7d\u5747\u8861\u4e09\u4e2a\u7ef4\u5ea6\u63d0\u4f9b\u4e86\u7ecf\u8fc7\u5927\u89c4\u6a21\u751f\u4ea7\u9a8c\u8bc1\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>SIGCOMM &#8217;24: ACM SIGCOMM 2024 Conference, August 4\u20138, 2024, Sydney, NSW, Australia https:\/\/dl.acm.org\/doi\/10.1145\/3651890.3672265 \u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a\uff1a\u4f20\u7edf\u6570\u636e\u4e2d\u5fc3\u7f51\u7edc\u4e3a\u4f55\u4e0d\u9002\u5408 LLM \u8bad\u7ec3 \u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bad\u7ec3\u5bf9\u6570\u636e\u4e2d\u5fc3\u7f51\u7edc\u63d0\u51fa\u4e86\u5168\u65b0\u7684\u6311\u6218\uff0c\u4f20\u7edf\u9762\u5411\u901a\u7528\u4e91\u8ba1\u7b97\u8bbe\u8ba1\u7684\u7f51\u7edc\u67b6\u6784\u5df2\u4e0d\u518d\u9002\u7528\u3002\u963f\u91cc\u5df4\u5df4\u5728\u5b9e\u9645\u751f\u4ea7\u4e2d\u89c2\u5bdf\u5230\u4e24\u4e2a\u6839\u672c\u6027\u7684\u4e0d\u5339\u914d\u3002 \u7b2c\u4e00\u4e2a &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1181\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":6,"featured_media":1182,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23],"tags":[],"class_list":["post-1181","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1181","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\/6"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1181"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1181\/revisions"}],"predecessor-version":[{"id":1184,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1181\/revisions\/1184"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1182"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1181"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1181"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1181"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}