\u9ad8\u6548\u80fd AI \u7db2\u8def\u8981\u6c42\u5c0d\u6240\u6709\u4ea4\u63db\u7bc0\u9ede\u9032\u884c\u6301\u7e8c\u7684\u6d41\u91cf\u76e3\u63a7\u3002\u5728\u5206\u6563\u5f0f\u8a13\u7df4\u904e\u7a0b\u4e2d\uff0c\u540c\u6b65\u5f0f\u96c6\u5408\u901a\u8a0a\u64cd\u4f5c\u6703\u7d66\u53e2\u96c6 Fabric \u5e36\u4f86\u5de8\u5927\u58d3\u529b\u3002<\/p>\n

\u5fae\u7a81\u767c\u3001Incast \u8207 PFC \u66ab\u505c\u98a8\u66b4<\/h3>\n

\u540c\u6b65\u5f0f\u96c6\u5408\u901a\u8a0a\u64cd\u4f5c\u6703\u5728\u5f8c\u7aef Fabric \u4e2d\u5f15\u767c\u7a81\u767c\u6d41\u91cf\u3002\u5728\u4e00\u6b21\u672c\u5730 AllToAll \u64cd\u4f5c\u4e2d\uff0c\u591a\u500b\u5de5\u4f5c\u7bc0\u9ede\u6703\u540c\u6642\u5411\u540c\u4e00\u500b\u63a5\u6536\u7bc0\u9ede\u50b3\u9001\u8cc7\u6599\u3002\u9019\u7a2e Incast \u6a21\u5f0f\u6703\u5c0e\u81f4\u7de9\u885d\u5340\u6025\u5287\u5806\u7a4d\uff0c\u4e26\u5e36\u4f86\u56b4\u91cd\u5c01\u5305\u907a\u5931\u3002<\/p>\n

\u7576\u7de9\u885d\u4f47\u5217\u88ab\u586b\u6eff\u6642\uff0c\u63a5\u6536\u7aef\u4ea4\u63db\u5668\u6703\u5411\u4e0a\u6e38\u50b3\u9001\u512a\u5148\u7d1a\u6d41\u91cf\u63a7\u5236\uff08PFC\uff09\u66ab\u505c\u5e40\uff0c\u4ee5\u9632\u6b62\u4e1f\u5305\u3002\u9019\u7a2e\u53cd\u58d3\u6703\u4f7f\u7db2\u8def\u4e2d\u8a72\u512a\u5148\u7d1a\u4f47\u5217\u4e0a\u7684\u6240\u6709\u6d41\u91cf\u5168\u90e8\u66ab\u505c\u3002<\/p>\n

\n
• \u4e0a\u6e38\u4ea4\u63db\u5668\u6703\u5728\u7279\u5b9a\u512a\u5148\u7d1a\u4f47\u5217\u4e0a\u63a5\u6536\u5230 PFC \u66ab\u505c\u8a0a\u865f\u3002<\/li>\n
• \u4ea4\u63db\u5668\u6703\u66ab\u505c\u6574\u500b\u4f47\u5217\uff0c\u4f7f\u58c5\u585e\u6d41\u8207\u672a\u58c5\u585e\u7684\u53d7\u5bb3\u6d41\u540c\u6642\u505c\u6b62\u3002<\/li>\n
• \u9019\u7a2e\u53cd\u58d3\u6703\u8de8\u8d8a\u591a\u500b\u8df3\u9ede\u50b3\u64ad\uff0c\u9020\u6210\u968a\u9996\u963b\u585e\u8207\u56b4\u91cd\u7684\u901f\u7387\u632f\u76ea\u3002<\/li>\n<\/ul>\n

  \u4e0d\u7a69\u5b9a\u7684\u66ab\u505c\u5e40\u50b3\u64ad\u6703\u5728 Fabric \u4e2d\u5f62\u6210\u58c5\u585e\u6a39\u3002\u9019\u4e00\u904e\u7a0b\u6703\u589e\u52a0\u5c3e\u5ef6\u9072\uff0c\u4e26\u964d\u4f4e\u6574\u500b\u53e2\u96c6\u7684\u7e3d\u9ad4\u541e\u5410\u91cf\u3002<\/p>\n

  CLOS \u8207 RAIL \u67b6\u69cb\u4e2d\u7684\u62d3\u64b2\u74f6\u9838<\/h3>\n

  \u7db2\u8def\u67b6\u69cb\u6c7a\u7b56\u6703\u76f4\u63a5\u5f71\u97ff\u4f60\u7684 Fabric \u5728\u9762\u5c0d\u9ad8\u5f37\u5ea6 AI \u5de5\u4f5c\u8ca0\u8f09\u6642\u7684\u627f\u8f09\u80fd\u529b\u3002\u50b3\u7d71 CLOS \u7db2\u8def\u4f9d\u8cf4\u7b49\u50f9\u591a\u8def\u5f91\u8def\u7531\u4f86\u5206\u6563\u52d5\u614b\u6d41\u91cf\u9700\u6c42\u3002<\/p>\n

  \u7136\u800c\uff0c\u6a19\u6e96 CLOS \u8a2d\u8a08\u5728 all-reduce \u64cd\u4f5c\u671f\u9593\uff0c\u5f80\u5f80\u96e3\u4ee5\u61c9\u5c0d\u975c\u614b\u6d41\u7d81\u5b9a\u8207\u8ca0\u8f09\u5931\u8861\u554f\u984c\u3002<\/p>\n

  \n\n\n\n\n<\/colgroup>\n\n\n\n\n

  \u7db2\u8def\u8a2d\u8a08<\/th>\n	\u6838\u5fc3\u6d41\u91cf\u5047\u8a2d<\/th>\n	\u4e3b\u8981\u7db2\u8def\u9650\u5236<\/th>\n<\/tr>\n
  CLOS \u67b6\u69cb<\/td>\n	\u5047\u8a2d\u6d41\u91cf\u7531\u5f7c\u6b64\u7368\u7acb\u3001\u751f\u547d\u9031\u671f\u8f03\u77ed\u7684\u5fae\u670d\u52d9\u6d41\u7d44\u6210<\/td>\n	\u6a19\u6e96 ECMP \u8def\u7531\u6703\u5c0e\u81f4\u6d41\u78b0\u649e<\/td>\n<\/tr>\n
  RAIL \u6700\u4f73\u5316\u67b6\u69cb<\/td>\n	\u5c08\u70ba\u540c\u6b65\u3001\u9577\u6301\u7e8c\u6642\u9593\u7684\u5927\u8c61\u6d41\u800c\u8a2d\u8a08<\/td>\n	\u8d85\u984d\u8a02\u95b1\u6bd4\u6703\u9020\u6210\u74f6\u9838<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n GPU \u53e2\u96c6\u4e2d\u7684\u6b21\u4f73\u4f5c\u696d\u653e\u7f6e\u9084\u6703\u9032\u4e00\u6b65\u60e1\u5316\u6574\u9ad4\u7db2\u8def\u6548\u80fd\u3002\u7cdf\u7cd5\u7684\u6392\u7a0b\u5668\u6c7a\u7b56\u6703\u89f8\u767c\u5c40\u90e8\u8449\u4ea4\u63db\u5668\u74f6\u9838\uff0c\u5f9e\u800c\u589e\u52a0\u7e3d\u4f5c\u696d\u5b8c\u6210\u6642\u9593\u3002<\/p>\n \u9762\u5411 AI \u6a21\u578b\u8a13\u7df4\u7684\u9059\u6e2c\u7b56\u7565<\/strong><\/h2>\n \u73fe\u4ee3\u53e2\u96c6\u67b6\u69cb\u9700\u8981\u9ad8\u89e3\u6790\u5ea6\u6d41\u91cf\u9059\u6e2c\uff0c\u624d\u80fd\u89c0\u6e2c\u5f8c\u7aef Fabric \u7684\u6548\u80fd\u3002\u4f60\u5fc5\u9808\u76e3\u63a7\u5373\u6642\u7db2\u8def\u72c0\u614b\uff0c\u624d\u80fd\u7dad\u6301\u53ef\u9760\u7684 AI \u7db2\u8def\u74b0\u5883\u3002<\/p>\n \u5373\u6642\u9059\u6e2c\u8207\u5e36\u5167\u6d41\u91cf\u76e3\u63a7<\/h3>\n \u5728 AI \u6a21\u578b\u8a13\u7df4\u904e\u7a0b\u4e2d\uff0c\u4f60\u9700\u8981\u5c0d\u4ea4\u63db\u5668\u7de9\u885d\u5340\u8207\u8cc7\u6599\u5c01\u5305\u6a19\u8a18\u9032\u884c\u7cbe\u78ba\u8996\u89ba\u5316\u3002\u9ad8\u6548\u80fd\u4ea4\u63db\u786c\u9ad4\u85c9\u52a9\u786c\u9ad4\u52a0\u901f\u8655\u7406\uff0c\u5728\u5f8c\u7aef Fabric \u4e2d\u4e32\u6d41\u50b3\u8f38\u5373\u6642\u8cc7\u6599\u3002\u5148\u9032\u7684\u6536\u96c6\u5668\u7db2\u8def\u6703\u8ffd\u8e64\u7cfb\u7d71\u8a0a\u865f\uff0c\u4ee5\u7dad\u6301 Fabric \u5065\u5eb7\u72c0\u614b\u3002<\/p>\n \n\n\n\n<\/colgroup>\n\n\n\n\n\n\n KPI \u540d\u7a31<\/strong><\/th>\n \u8aaa\u660e\u8207\u7dad\u904b\u610f\u7fa9<\/strong><\/th>\n<\/tr>\n ECN \u6a19\u8a18\u6bd4\u7387<\/strong><\/td>\n \u7528\u65bc\u8861\u91cf\u88ab\u6a19\u8a18 CE \u78bc\u9ede\u7684\u8cc7\u6599\u5c01\u5305\u6bd4\u4f8b\uff0c\u4ee5\u6aa2\u6e2c\u6301\u7e8c\u6027\u58c5\u585e\u3002<\/td>\n<\/tr>\n PFC \u4e8b\u4ef6\u901f\u7387<\/strong><\/td>\n \u7528\u65bc\u8ffd\u8e64\u6d41\u63a7\u66ab\u505c\u8a0a\u865f\uff0c\u9632\u6b62\u4ea4\u63db\u5668\u4f47\u5217\u6ea2\u4f4d\u3002<\/td>\n<\/tr>\n \u7de9\u885d\u5340\u5360\u7528\u7387\uff08P99\uff09<\/strong><\/td>\n \u7528\u65bc\u64f7\u53d6 99 \u5206\u4f4d\u4f47\u5217\u5229\u7528\u7387\uff0c\u4ee5\u63ed\u793a\u6975\u7aef\u8a18\u61b6\u9ad4\u98fd\u548c\u60c5\u6cc1\u3002<\/td>\n<\/tr>\n \u91cd\u50b3\u7387<\/strong><\/td>\n \u7528\u65bc\u8b58\u5225\u786c\u9ad4\u93c8\u8def\u4e0a\u7684\u4e1f\u5305\u60c5\u6cc1\uff0c\u5f9e\u800c\u5b9a\u4f4d\u50b3\u8f38\u9000\u5316\u554f\u984c\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u4f60\u5fc5\u9808\u8ffd\u8e64\u95dc\u9375\u786c\u9ad4\u8a08\u6578\u5668\uff0c\u76e1\u65e9\u767c\u73fe\u6548\u80fd\u4e0b\u964d\u3002\u61c9\u59cb\u7d42\u512a\u5148\u95dc\u6ce8 Adp Retx \u8207 PCI \u8a08\u6578\u5668\uff0c\u4ee5\u5b9a\u4f4d\u5e38\u898b\u74f6\u9838\u3002\u6a19\u6e96\u7db2\u8def\u6846\u67b6\u6703\u5728\u908a\u7de3\u7aef\u90e8\u7f72\u81ea\u9069\u61c9\u53d6\u6a23\u8207\u58d3\u7e2e\u3002\u9019\u9805\u6280\u8853\u53ef\u5728\u4fdd\u6301\u5b8c\u6574\u7cbe\u5ea6\u7684\u540c\u6642\uff0c\u5c07\u9059\u6e2c\u8cc7\u6599\u7e3d\u91cf\u6700\u591a\u6e1b\u5c11 80%\u3002<\/p>\n \u9ad8\u89e3\u6790\u5ea6\u6536\u96c6\u5668\u6703\u64f7\u53d6\u7dad\u904b\u8cc7\u6599\uff0c\u4ee5\u6539\u9032\u6574\u9ad4\u7db2\u8def\u8a2d\u8a08\u3002\u5206\u6563\u5f0f\u6536\u96c6\u5668\u6703\u5206\u6790\u6d41\u91cf\u6d41\u5411\uff0c\u4fdd\u8b77\u5f8c\u7aef\u7db2\u8def\u7a69\u5b9a\u6027\u3002\u4f60\u61c9\u8a55\u4f30\u6d3b\u8e8d\u4ea4\u63db\u5668\u4e0a\u7684\u5ef6\u9072\u6307\u6a19\uff0c\u4ee5\u6e2c\u91cf\u8def\u5f91\u6642\u5ef6\u3002\u8ffd\u8e64\u5ef6\u9072\u6307\u6a19\u80fd\u70ba\u6301\u7e8c\u6700\u4f73\u5316\u63d0\u4f9b\u6e05\u6670\u7684\u7dad\u904b\u56de\u994b\u3002\u9ad8\u541e\u5410\u8981\u6c42\u6bcf\u4e00\u689d Fabric \u8def\u5f91\u90fd\u5177\u5099\u4f4e\u5ef6\u9072\u3002<\/p>\n \u9762\u5411\u58c5\u585e\u7684\u6642\u5e8f\u7570\u5e38\u6aa2\u6e2c<\/h3>\n \u4e3b\u52d5\u5f0f\u9059\u6e2c\u7cfb\u7d71\u53ef\u5728\u7de9\u885d\u4f47\u5217\u6ea2\u4f4d\u524d\u9632\u6b62\u7db2\u8def\u58c5\u585e\u3002\u4f60\u53ef\u4ee5\u90e8\u7f72 LSTM \u8207 1D-CNN \u7b49\u6df1\u5ea6\u5b78\u7fd2\u6642\u5e8f\u6a21\u578b\u4f86\u8655\u7406\u9023\u7e8c\u7684\u6307\u6a19\u6d41\u3002Autoformer\u3001Informer \u8207 TimesNet \u7b49\u5148\u9032 Transformer \u67b6\u69cb\u53ef\u5728\u52d5\u614b\u8a13\u7df4\u968e\u6bb5\u9810\u6e2c\u6d41\u91cf\u8da8\u52e2\u3002\u73fe\u4ee3\u7dad\u904b\u9ad4\u7cfb\u9084\u6703\u5229\u7528 MOMENT \u9019\u985e\u901a\u7528\u6642\u5e8f\u57fa\u790e\u6a21\u578b\u4f86\u6aa2\u6e2c\u7d30\u5fae\u7684\u904b\u884c\u7570\u5e38\u3002<\/p>\n \u4e00\u500b\u57fa\u65bc 18 \u500b\u6708\u6392\u7a0b\u6b77\u53f2\u8a13\u7df4\u800c\u6210\u7684\u6642\u5e8f\u6a21\u5f0f\u5f15\u64ce\uff0c\u53ef\u5728\u9700\u6c42\u9ad8\u5cf0\u771f\u6b63\u51fa\u73fe\u524d 2 \u5230 8 \u5206\u9418\u9810\u6e2c\u7b97\u529b\u9700\u6c42\uff0c\u4e26\u9810\u5148\u653e\u7f6e\u5f35\u91cf\u5206\u7247\u3002<\/p><\/blockquote>\n \u9019\u7a2e\u9810\u6e2c\u6027\u64cd\u4f5c\u53ef\u6d88\u9664\u5927\u578b\u53e2\u96c6\u4e2d\u7684\u7b97\u529b\u9592\u7f6e\uff0c\u56e0\u70ba\u9019\u4e9b\u53e2\u96c6\u6bcf\u9031\u6703\u56e0\u88ab\u52d5\u6392\u7a0b\u5ef6\u9072\u800c\u640d\u5931\u6578\u5343 GPU \u5c0f\u6642\u3002\u8a72\u9059\u6e2c\u6846\u67b6\u4ee5 100 \u6beb\u79d2\u9593\u9694\u6301\u7e8c\u63a1\u96c6\u8cc7\u6599\u3002<\/p>\n \n \u6bcf 100 \u6beb\u79d2\u63a1\u96c6\u4e00\u6b21 GPU \u6eab\u5ea6\u3001\u8a18\u61b6\u9ad4\u58d3\u529b\u3001NVLink \u5229\u7528\u7387\u8207\u4f5c\u696d\u4f47\u5217\u6df1\u5ea6\u3002<\/li>\n \u4f7f\u7528 Prometheus \u7684 rate()<\/code> \u51fd\u5f0f\u4f86\u8a08\u7b97\u91cd\u50b3\u7387\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u4f7f\u7528\u539f\u59cb\u8a08\u6578\u5668\u503c\u3002<\/li>\n \u5c07\u91cd\u50b3\u5cf0\u503c\u8207 NCCL \u6548\u80fd\u6307\u6a19\u76f4\u63a5\u95dc\u806f\uff0c\u4ee5\u78ba\u8a8d\u662f\u5426\u5b58\u5728\u4e1f\u5305\u3002<\/li>\n \u70ba\u9ad8\u5f80\u8fd4\u6642\u5ef6\u3001TLP \u932f\u8aa4\u8207\u93c8\u8def\u9000\u5316\u8a2d\u5b9a\u95dc\u9375\u8b66\u793a\u3002<\/li>\n<\/ul>\n \u4e3b\u52d5\u9810\u6e2c\u80fd\u5920\u5728\u91cd\u5ea6\u96c6\u5408\u901a\u8a0a\u671f\u9593\u7dad\u6301\u9ad8\u6548\u7684 AI \u7db2\u8def\u6a19\u6e96\u3002\u667a\u6167\u7570\u5e38\u6aa2\u6e2c\u53ef\u5c07\u5f8c\u7aef\u7db2\u8def\u8f49\u5316\u70ba\u5177\u5099\u97cc\u6027\u7684 Fabric\u3002\u59a5\u5584\u7684\u8cc7\u6599\u651d\u53d6\u80fd\u5920\u5f37\u5316\u5927\u578b\u7b97\u529b\u53e2\u96c6\u4e2d\u7684\u7a69\u5065\u7db2\u8def\u7dad\u904b\u3002\u667a\u6167\u9059\u6e2c\u7b56\u7565\u900f\u904e\u4fdd\u8b77\u6d3b\u8e8d\u4f5c\u696d\u514d\u53d7\u7de9\u885d\u5340\u8017\u76e1\u5f71\u97ff\uff0c\u589e\u5f37\u4f01\u696d\u7d1a AI \u7db2\u8def\u67b6\u69cb\u3002\u7d30\u7dfb\u7684\u76e3\u63a7\u53ef\u5728\u9ad8\u8ca0\u8f09\u4e0b\u4fdd\u969c\u7a69\u5b9a\u7684\u7db2\u8def\u6548\u80fd\u3002<\/p>\n \u91dd\u5c0d AI \u8a13\u7df4\u5de5\u4f5c\u8ca0\u8f09\u6700\u4f73\u5316 Fabric<\/strong><\/h2>\n \u4f60\u53ef\u4ee5\u900f\u904e\u6821\u6e96\u50b3\u8f38\u5354\u5b9a\u4e26\u6700\u4f73\u5316\u5f8c\u7aef\u57fa\u790e\u8a2d\u65bd\u4e2d\u7684\u4f5c\u696d\u6392\u7a0b\u4f86\u6d88\u9664 Fabric \u74f6\u9838\u3002\u660e\u667a\u7684\u5de5\u7a0b\u6c7a\u7b56\u6703\u76f4\u63a5\u63d0\u5347\u5927\u898f\u6a21\u53e2\u96c6\u904b\u884c\u671f\u9593\u7684\u7db2\u8def\u6548\u80fd\u3002<\/p>\n \u50b3\u8f38\u5354\u5b9a\u8207 DCQCN \u53c3\u6578\u8abf\u6821<\/h3>\n \u8cc7\u6599\u4e2d\u5fc3\u91cf\u5316\u58c5\u585e\u901a\u77e5\uff08DCQCN\uff09\u7528\u65bc\u63a7\u5236 RoCEv2 \u7db2\u8def\u4e2d\u7684\u58c5\u585e\u3002\u5728\u9ad8\u5f37\u5ea6\u64cd\u4f5c\u671f\u9593\uff0c\u4f60\u5fc5\u9808\u8abf\u6574 DCQCN \u53c3\u6578\uff0c\u4ee5\u5728\u7cfb\u7d71\u541e\u5410\u8207\u8cc7\u6599\u5c01\u5305\u6642\u5ef6\u4e4b\u9593\u53d6\u5f97\u5e73\u8861\u3002\u7576\u4f47\u5217\u6df1\u5ea6\u8d85\u904e\u95be\u503c\u9650\u5236\u6642\uff0c\u4ea4\u63db\u786c\u9ad4\u6703\u4f7f\u7528\u660e\u78ba\u58c5\u585e\u901a\u77e5\uff08ECN\uff09\u5c0d\u8cc7\u6599\u5c01\u5305\u9032\u884c\u6a19\u8a18\u3002<\/p>\n \n \u6839\u64da\u93c8\u8def\u983b\u5bec\u8207\u57e0\u4f47\u5217\u5bb9\u91cf\u8a2d\u5b9a min_kt<\/code> \u8207 max_kt<\/code> \u7de9\u885d\u95be\u503c\u3002<\/li>\n \u8a2d\u5b9a\u76ee\u6a19\u901f\u7387\u4e0b\u964d\u53c3\u6578 ai<\/code>\uff0c\u4ee5\u9632\u6b62\u5728\u5fae\u7a81\u767c\u671f\u9593\u767c\u751f\u56b4\u91cd\u6d41\u91cf\u9650\u901f\u3002<\/li>\n \u8abf\u6574\u8a08\u6642\u5668\u53c3\u6578 hai<\/code>\uff0c\u4ee5\u4fbf\u5728\u58c5\u585e\u89e3\u9664\u5f8c\u5feb\u901f\u6062\u5fa9\u76ee\u6a19\u983b\u5bec\u3002<\/li>\n<\/ul>\n \u932f\u8aa4\u7684\u53c3\u6578\u503c\u6703\u5728\u9ad8\u6548\u80fd AI Fabric \u4e2d\u5f15\u767c\u901f\u7387\u632f\u76ea\u3002\u7cbe\u7d30\u8abf\u6821\u5354\u5b9a\u8a2d\u5b9a\u80fd\u5920\u70ba\u540c\u6b65\u5f0f GPU \u901a\u8a0a\u5efa\u7acb\u53ef\u9760\u7684 AI \u7db2\u8def\u3002\u59a5\u7576\u7684\u53c3\u6578\u9078\u64c7\u53ef\u4ee5\u5728\u9ad8\u6d41\u91cf\u7a81\u767c\u4e0b\u7a69\u5b9a\u6838\u5fc3\u7db2\u8def\u57fa\u790e\u8a2d\u65bd\u3002\u6700\u4f73\u5316\u5f8c\u7684\u50b3\u8f38\u63a7\u5236\u53ef\u964d\u4f4e\u53e2\u96c6\u4e2d\u6bcf\u500b\u7bc0\u9ede\u7684\u5c3e\u5ef6\u9072\u3002\u73fe\u4ee3\u9ad8\u6548\u80fd\u7db2\u8def\u53ef\u78ba\u4fdd\u8907\u96dc\u96c6\u5408\u901a\u8a0a\u6b65\u9a5f\u4e2d\u7684\u5e73\u9806\u50b3\u8f38\u3002<\/p>\n \u81ea\u9069\u61c9\u8def\u7531\u8207\u4f5c\u696d\u653e\u7f6e\u6700\u4f73\u5316<\/h3>\n \u975c\u614b ECMP \u8def\u7531\u6703\u5728 all-to-all \u901a\u8a0a\u671f\u9593\u9020\u6210\u8cc7\u6599\u5c01\u5305\u78b0\u649e\u3002\u81ea\u9069\u61c9\u8def\u7531\u786c\u9ad4\u6703\u52d5\u614b\u8a55\u4f30\u93c8\u8def\u5229\u7528\u7387\u3002\u4ea4\u63db\u5668\u6703\u5c07\u8cc7\u6599\u5c01\u5305\u7051\u9001\u5230\u53ef\u7528\u8def\u5f91\u4e0a\uff0c\u4ee5\u7dad\u6301\u6700\u4f73 Fabric \u6d41\u52d5\u3002\u9019\u7a2e\u52d5\u614b\u5206\u914d\u80fd\u5920\u63d0\u5347\u591a\u5c64\u62d3\u64b2\u4e2d\u7684\u6574\u9ad4\u7db2\u8def\u6548\u7387\u3002<\/p>\n \u62d3\u64b2\u611f\u77e5\u4f5c\u696d\u6392\u7a0b\u53ef\u900f\u904e\u5c07\u76f8\u95dc\u8a08\u7b97\u4efb\u52d9\u96c6\u4e2d\u653e\u7f6e\uff0c\u907f\u514d\u93c8\u8def\u98fd\u548c\u3002\u59a5\u5584\u7684\u653e\u7f6e\u7b56\u7565\u53ef\u4fdd\u8b77\u4f60\u7684\u5f8c\u7aef\u7db2\u8def\u74b0\u5883\u514d\u53d7\u5c40\u90e8\u58c5\u585e\u5f71\u97ff\u3002\u4e00\u9805\u5b9a\u91cf\u6a21\u64ec\u7814\u7a76\u8a55\u4f30\u4e86\u4e0d\u540c\u653e\u7f6e\u7b56\u7565\u7684\u7dad\u904b\u6536\u76ca\u3002<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n\n \u6539\u9032\u9762\u5411<\/th>\n \u5b9a\u91cf\u6a21\u64ec\u7814\u7a76\u4e2d\u7684\u8b49\u64da<\/th>\n \u91cf\u5316\u652f\u6490<\/th>\n<\/tr>\n \u63d0\u5347\u7db2\u8def\u5229\u7528\u7387 \/ \u6e1b\u5c11\u788e\u7247\u5316<\/strong><\/td>\n \u7814\u7a76\u986f\u793a\uff0c\u6392\u7a0b\u5668\u80fd\u5920\u6709\u7b56\u7565\u5730\u5c07\u5c0f\u578b\u4f5c\u696d\u653e\u7f6e\u5728 NVLink \u7db2\u57df\u4e2d\u7684\u7279\u5b9a\u7bc0\u9ede\u4e0a\u3002<\/td>\n \u71b1\u529b\u5716\u986f\u793a\uff0c\u5c0f\u4f5c\u696d\u88ab\u96c6\u4e2d\u653e\u7f6e\u5728\u6bcf\u500b\u7db2\u57df\u6700\u5f8c\u5169\u500b\u7bc0\u9ede\u4e0a\uff0c\u4ee5\u76e1\u91cf\u6e1b\u5c11\u9ad8\u983b\u5bec\u7db2\u8def\u8cc7\u6e90\u788e\u7247\u5316\uff0c\u5f9e\u800c\u70ba\u66f4\u5927\u4f5c\u696d\u4fdd\u7559\u8cc7\u6e90\u3002<\/td>\n<\/tr>\n \u5c07\u6548\u80fd\u958b\u92b7\u964d\u81f3\u6700\u4f4e<\/strong><\/td>\n \u7814\u7a76\u5c07\u9019\u7a2e\u62d3\u64b2\u611f\u77e5\u6392\u7a0b\u5668\u7684\u5360\u7528\u7387\uff08\u53ef\u8996\u70ba\u5229\u7528\u7387\u4ee3\u7406\u6307\u6a19\uff09\u8207\u7406\u8ad6\u4e0a\u7684\u6700\u512a\u3001\u4f46\u4e0d\u5177\u5099\u62d3\u64b2\u611f\u77e5\u80fd\u529b\u7684\u6392\u7a0b\u5668\u9032\u884c\u6bd4\u8f03\u3002<\/td>\n \u7d50\u679c\u8868\u660e\uff0c\u5360\u7528\u7387\u5dee\u8ddd\u50c5\u7d04\u70ba 1%\uff08\u5716 5\uff09\uff0c\u8b49\u660e\u5373\u4f7f\u907f\u514d\u4e86\u7cdf\u7cd5\u653e\u7f6e\u5e36\u4f86\u7684\u6548\u80fd\u640d\u5931\uff0c\u4f9d\u7136\u80fd\u5920\u7dad\u6301\u5f88\u9ad8\u7684\u5229\u7528\u7387\u3002<\/td>\n<\/tr>\n \u652f\u6490\u7814\u7a76\u7684\u898f\u6a21<\/strong><\/td>\n \u76f8\u95dc\u8b49\u64da\u57fa\u65bc\u4e00\u500b\u7d30\u7c92\u5ea6\u3001\u6642\u9593\u52a0\u901f\u7684 Slurm \u6a21\u64ec\u6846\u67b6\u3002<\/td>\n \u8a72\u6a21\u64ec\u5c0d\u4e00\u500b 5,000 \u7bc0\u9ede\u53e2\u96c6\u9032\u884c\u5efa\u6a21\uff0c\u5728 7 \u5929\u6642\u9593\u5167\u56de\u653e\u4e86 15,000 \u500b\u4f5c\u696d\uff0c\u56e0\u6b64\u5177\u5099\u7d71\u8a08\u610f\u7fa9\u4e0a\u7684\u986f\u8457\u6027\u8207\u91cf\u5316\u50f9\u503c\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u667a\u6167\u653e\u7f6e\u6f14\u7b97\u6cd5\u53ef\u4ee5\u9694\u96e2\u9ad8\u5f37\u5ea6\u5de5\u4f5c\u8ca0\u8f09\uff0c\u4f7f\u8a08\u7b97\u6838\u5fc3\u4fdd\u6301\u5145\u5206\u6d3b\u8e8d\u3002\u7576\u6392\u7a0b\u5668\u80fd\u8b93\u4f5c\u696d\u9700\u6c42\u8207\u5be6\u9ad4\u93c8\u8def\u5bb9\u91cf\u76f8\u5339\u914d\u6642\uff0c\u6574\u9ad4 AI \u5de5\u4f5c\u8ca0\u8f09\u57f7\u884c\u6703\u66f4\u5feb\u3002\u6bcf\u4e00\u7a2e\u73fe\u4ee3 AI \u5de5\u4f5c\u8ca0\u8f09\u90fd\u9700\u8981\u7cbe\u78ba\u653e\u7f6e\uff0c\u4ee5\u6700\u5927\u5316\u786c\u9ad4\u56de\u5831\u3002\u9ad8\u6548\u7684\u4f5c\u696d\u5206\u914d\u80fd\u5920\u5f37\u5316\u73fe\u4ee3\u5206\u6563\u5f0f\u74b0\u5883\u4e2d\u7684\u7a69\u5065 AI \u7db2\u8def\u3002\u5148\u9032\u6392\u7a0b\u53ef\u4ee5\u4fdd\u8b77\u4f01\u696d\u7d1a AI \u7db2\u8def\u90e8\u7f72\uff0c\u907f\u514d\u7a81\u767c\u901a\u8a0a\u964d\u901f\u3002\u900f\u904e\u5c07\u901a\u8a0a\u6a21\u5f0f\u8207\u4ea4\u63db\u9023\u7dda\u95dc\u4fc2\u76f8\u5339\u914d\uff0c\u4f60\u53ef\u4ee5\u4fdd\u7559\u66f4\u591a Fabric \u5bb9\u91cf\u3002\u660e\u667a\u7684\u6392\u7a0b\u6c7a\u7b56\u80fd\u5920\u589e\u5f37\u4f01\u696d\u8cc7\u6599\u4e2d\u5fc3\u4e2d\u7684\u73fe\u4ee3 AI \u7db2\u8def\u7dad\u904b\u3002\u900f\u904e\u5c07\u81ea\u9069\u61c9\u8def\u7531\u8207\u62d3\u64b2\u611f\u77e5\u6392\u7a0b\u7d50\u5408\u61c9\u7528\u65bc\u6240\u6709 AI \u8a13\u7df4\u5de5\u4f5c\u8ca0\u8f09\uff0c\u4f60\u53ef\u4ee5\u5be6\u73fe\u7a69\u5b9a\u7684\u53e2\u96c6\u541e\u5410\u3002\u6709\u6548\u7684 Fabric \u8abf\u6821\u53ef\u4f7f\u57fa\u790e\u8a2d\u65bd\u66f4\u9069\u914d\u9ad8\u5f37\u5ea6 AI \u8a13\u7df4\u5de5\u4f5c\u8ca0\u8f09\u3002<\/p>\n \n <\/div>\n<\/div>\n \u900f\u904e\u5c07\u6301\u7e8c\u9059\u6e2c\u3001\u81ea\u9069\u61c9\u8def\u7531\u8207\u7d93\u904e\u8abf\u6821\u7684\u50b3\u8f38\u53c3\u6578\u6574\u5408\u9032\u4e00\u500b\u4e3b\u52d5\u56de\u994b\u9589\u74b0\uff0c\u4f60\u53ef\u4ee5\u6700\u5927\u5316 GPU \u7b97\u529b\u5229\u7528\u7387\u3002\u4e3b\u52d5\u5f0f\u58c5\u585e\u7ba1\u7406\u8207\u62d3\u64b2\u611f\u77e5\u4f5c\u696d\u653e\u7f6e\u80fd\u5920\u6d88\u9664\u7db2\u8def\u58c5\u585e\u3002\u9019\u4e00\u7b56\u7565\u53ef\u63d0\u5347\u5f8c\u7aef\u53e2\u96c6\u7684\u6574\u9ad4\u8a13\u7df4\u6548\u7387\u3002\u73fe\u4ee3 AI \u7db2\u8def\u900f\u904e\u5e73\u8861\u6d41\u91cf\u9700\u6c42\uff0c\u5728\u9ad8\u5f37\u5ea6\u64cd\u4f5c\u671f\u9593\u7dad\u6301\u9ad8\u7db2\u8def\u6548\u80fd\u4e26\u76e1\u53ef\u80fd\u964d\u4f4e\u6642\u5ef6\u3002<\/p>\n \u7db2\u8def\u67b6\u69cb\u5e2b\u5fc5\u9808\u5b9a\u671f\u5b8c\u6210\u4e09\u9805 Fabric \u7a3d\u6838\uff1a<\/p>\n \n \u6aa2\u67e5\u4ea4\u63db\u5668\u7de9\u885d\u95be\u503c\uff0c\u9632\u6b62\u4e1f\u5305\u3002<\/li>\n \u6821\u6e96 ECN \u6a19\u8a18\u901f\u7387\uff0c\u63a7\u5236\u4f47\u5217\u6210\u9577\u3002<\/li>\n \u66f4\u65b0\u8def\u7531\u53c3\u6578\uff0c\u63d0\u5347\u7db2\u8def\u6548\u7387\u3002<\/li>\n<\/ul>\n \u9ad8\u6548\u7684 AI \u7db2\u8def\u80fd\u5920\u5efa\u7acb\u5177\u5099\u97cc\u6027\u7684\u5f8c\u7aef Fabric\u3002\u667a\u6167\u7db2\u8def\u5be6\u52d9\u53ef\u4fdd\u8b77\u5f8c\u7aef\u7bc0\u9ede\u514d\u53d7\u56b4\u91cd\u74f6\u9838\u5f71\u97ff\u3002\u59a5\u5584\u7684\u7db2\u8def\u8a2d\u8a08\u80fd\u5920\u6700\u5927\u5316\u53e2\u96c6\u7522\u51fa\u3002<\/p>\n \u5e38\u898b\u554f\u984c<\/strong><\/h2>\n \u5728 AI \u6a21\u578b\u8a13\u7df4\u671f\u9593\uff0c\u662f\u4ec0\u9ebc\u5c0e\u81f4\u4e86 PFC \u66ab\u505c\u98a8\u66b4\uff1f<\/h3>\n \u540c\u6b65\u5f0f\u96c6\u5408\u901a\u8a0a\u64cd\u4f5c\u6703\u7522\u751f\u6d41\u91cf\u5fae\u7a81\u767c\u8207\u56b4\u91cd\u7684 Incast\u3002\u7576\u4ea4\u63db\u5668\u4f47\u5217\u88ab\u586b\u6eff\u6642\uff0c\u63a5\u6536\u7aef\u4ea4\u63db\u5668\u6703\u5411\u4e0a\u6e38\u50b3\u9001 PFC \u66ab\u505c\u5e40\u3002\u9019\u4e9b\u66ab\u505c\u8a0a\u865f\u6703\u8de8\u8d8a\u591a\u500b\u8df3\u9ede\u50b3\u64ad\uff0c\u5728\u4f60\u7684\u5f8c\u7aef Fabric \u4e2d\u9020\u6210\u53cd\u58d3\u3001\u968a\u9996\u963b\u585e\u8207\u66ab\u505c\u98a8\u66b4\u3002<\/p>\n \u660e\u78ba\u58c5\u585e\u901a\u77e5\u5982\u4f55\u7de9\u89e3 Fabric \u58c5\u585e\uff1f<\/h3>\n \u7576\u4ea4\u63db\u5668\u4f47\u5217\u6df1\u5ea6\u8d85\u904e\u4f60\u8a2d\u5b9a\u7684\u95be\u503c\u6642\uff0c\u660e\u78ba\u58c5\u585e\u901a\u77e5\u6703\u5c0d\u8cc7\u6599\u5c01\u5305\u6a19\u982d\u9032\u884c\u6a19\u8a18\u3002\u8a72\u8a0a\u865f\u80fd\u5920\u53ca\u65e9\u901a\u77e5\u7aef\u4e3b\u6a5f\u5b58\u5728\u6301\u7e8c\u6027\u58c5\u585e\u3002\u56e0\u6b64\uff0c\u7aef\u4e3b\u6a5f\u50b3\u8f38\u5354\u5b9a\u6703\u5728\u4e1f\u5305\u767c\u751f\u8207\u7de9\u885d\u4f47\u5217\u6ea2\u4f4d\u4e4b\u524d\u5148\u884c\u964d\u4f4e\u50b3\u9001\u901f\u7387\u3002<\/p>\n \u70ba\u4ec0\u9ebc\u5728\u5f8c\u7aef\u7db2\u8def\u4e2d\u61c9\u512a\u5148\u9078\u64c7\u81ea\u9069\u61c9\u8def\u7531\uff0c\u800c\u4e0d\u662f\u975c\u614b ECMP\uff1f<\/h3>\n \u975c\u614b ECMP \u6703\u6cbf\u56fa\u5b9a\u8def\u5f91\u8f49\u9001\u6d41\u91cf\uff0c\u5728\u540c\u6b65\u5f0f\u5927\u8c61\u6d41\u5834\u666f\u4e0b\u5f15\u767c\u6d41\u78b0\u649e\u3002\u81ea\u9069\u61c9\u8def\u7531\u5247\u6703\u52d5\u614b\u8a55\u4f30\u5373\u6642\u93c8\u8def\u5229\u7528\u7387\u3002\u4ea4\u63db\u5668\u6703\u5c07\u8cc7\u6599\u5c01\u5305\u7051\u9001\u5230\u6240\u6709\u53ef\u7528\u7684\u904b\u884c\u8def\u5f91\u4e0a\uff0c\u5f9e\u800c\u6700\u5927\u5316\u6574\u9ad4\u7db2\u8def\u541e\u5410\uff0c\u4e26\u9632\u6b62\u5c40\u90e8\u983b\u5bec\u74f6\u9838\u3002<\/p>\n \u62d3\u64b2\u611f\u77e5\u4f5c\u696d\u653e\u7f6e\u5982\u4f55\u63d0\u5347\u53e2\u96c6\u6548\u7387\uff1f<\/h3>\n \u62d3\u64b2\u611f\u77e5\u6392\u7a0b\u5668\u6703\u5c07\u76f8\u4e92\u95dc\u806f\u7684 GPU \u4efb\u52d9\u5206\u914d\u5230\u5171\u4eab NVLink \u7db2\u57df\u5167\u5f7c\u6b64\u9130\u8fd1\u7684\u5be6\u9ad4\u7bc0\u9ede\u4e0a\u3002\u9019\u7a2e\u667a\u6167\u653e\u7f6e\u65b9\u5f0f\u80fd\u5920\u9694\u96e2\u9ad8\u901a\u8a0a\u5f37\u5ea6\u6d41\u91cf\uff0c\u6e1b\u5c11\u591a\u5c64\u62d3\u64b2\u4e2d\u7684\u93c8\u8def\u788e\u7247\u5316\uff0c\u4e26\u9632\u6b62\u5171\u4eab\u8449\u7bc0\u9ede\u93c8\u8def\u4e0a\u7684\u983b\u5bec\u98fd\u548c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u5c11\u91cf\u7570\u5e38\u5de5\u4f5c\u8ca0\u8f09\u5c31\u53ef\u80fd\u5728\u591a\u7bc0\u9ede GPU \u53e2\u96c6\u4e2d\u9020\u6210\u5c3e\u5ef6\u9072\u3002\u9019\u7a2e\u5ef6\u9072\u6703\u62d6\u6162\u5206\u6563\u5f0f\u64cd\u4f5c\u4e2d\u7684\u4f5c\u696d\u5b8c\u6210\u6642\u9593\u3002\u4e8b\u5be6\u4e0a\uff0c\u7db2 […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":33970,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4815],"tags":[12709,12710,12711,4849],"class_list":["post-33973","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-america-dedicated-server-tc","tag-ai-cluster-network-congestion","tag-real-time-flow-telemetry","tag-gpu-cluster-optimization","tag-us-hosting-tc"],"acf":[],"yoast_head":"\n