{"id":5745,"date":"2025-07-30T10:09:31","date_gmt":"2025-07-30T02:09:31","guid":{"rendered":"https:\/\/www.edge-core.com\/?p=5745"},"modified":"2025-08-21T17:22:35","modified_gmt":"2025-08-21T09:22:35","slug":"powering-ai-at-scale-the-indispensable-role-of-data-center-bridging-with-pfc-and-ecn-for-large-gpu-deployments","status":"publish","type":"post","link":"https:\/\/www.edge-core.com\/cn\/blogs\/powering-ai-at-scale-the-indispensable-role-of-data-center-bridging-with-pfc-and-ecn-for-large-gpu-deployments\/","title":{"rendered":"\u5927\u89c4\u6a21 AI \u8d4b\u80fd\uff1a\u6570\u636e\u4e2d\u5fc3\u6865\u63a5\u6280\u672f\u5728\u5927\u89c4\u6a21 GPU \u90e8\u7f72\u4e2d\u53d1\u6325\u4e0d\u53ef\u6216\u7f3a\u7684\u4f5c\u7528"},"content":{"rendered":"

<\/div><\/div>
<\/span><\/div><\/div><\/div>
\"Mark<\/span><\/div><\/div><\/div>

\u4f5c\u8005\uff1a\u9a6c\u514b\u00b7\u54c8\u91cc\u65af <\/i><\/a><\/p>

\u53d1\u5e03\u4e8e 2025 \u5e74 7 \u6708 30 \u65e5<\/span><\/p>\n<\/div><\/div><\/div><\/div><\/div><\/div>

\u5bf9\u66f4\u667a\u80fd\u7684\u4eba\u5de5\u667a\u80fd\u6a21\u578b\uff08\u5c24\u5176\u662f\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u548c\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5\uff09\u7684\u4e0d\u61c8\u8ffd\u6c42\uff0c\u63a8\u52a8\u4e86\u5bf9\u8ba1\u7b97\u80fd\u529b\u524d\u6240\u672a\u6709\u7684\u9700\u6c42\u3002\u8fd9\u573a\u9769\u547d\u7684\u6838\u5fc3\u662f\u56fe\u5f62\u5904\u7406\u5355\u5143 (GPU)\uff0c\u5b83\u51ed\u501f\u5176\u5e76\u884c\u5904\u7406\u80fd\u529b\uff0c\u975e\u5e38\u9002\u5408\u5b9a\u4e49\u4eba\u5de5\u667a\u80fd\u8bad\u7ec3\u7684\u5bc6\u96c6\u578b\u8fed\u4ee3\u8ba1\u7b97\u3002\u7136\u800c\uff0c\u5728\u6570\u636e\u4e2d\u5fc3\u73af\u5883\u4e2d\u5145\u5206\u91ca\u653e\u6570\u5343\u4e2a\u4e92\u8fde GPU \u7684\u6f5c\u529b\u5e76\u975e\u4ec5\u4ec5\u662f\u7b80\u5355\u5730\u5c06\u5b83\u4eec\u63a5\u5165\u7f51\u7edc\u90a3\u4e48\u7b80\u5355\u3002\u5b83\u9700\u8981\u4e00\u4e2a\u80fd\u591f\u4ee5\u6700\u5c0f\u5ef6\u8fdf\u548c\u81f3\u5173\u91cd\u8981\u7684\u96f6\u4e22\u5305\u7387\u5904\u7406\u6d77\u91cf\u6570\u636e\u6d41\u7684\u590d\u6742\u7f51\u7edc\u57fa\u7840\u8bbe\u65bd\u3002\u8fd9\u6b63\u662f \u6570\u636e\u4e2d\u5fc3\u6865\u63a5 (DCB)<\/strong>\uff0c\u518d\u52a0\u4e0a\u5148\u8fdb\u7684\u6d41\u91cf\u63a7\u5236\u673a\u5236\uff0c\u5982 \u57fa\u4e8e\u4f18\u5148\u7ea7\u7684\u6d41\u91cf\u63a7\u5236\uff08PFC\uff09<\/strong> \u548c \u663e\u5f0f\u62e5\u585e\u901a\u77e5 (ECN)<\/strong>\uff0c\u53d8\u5f97\u7edd\u5bf9\u5fc5\u8981\u3002<\/p>\n

\u5927\u89c4\u6a21 GPU \u96c6\u7fa4\u4e2d\u7684\u7f51\u7edc\u62e5\u585e\u6982\u8ff0<\/h1>\n

\u60f3\u8c61\u4e00\u4e0b\uff0c\u4e00\u9879 AI \u8bad\u7ec3\u4efb\u52a1\u5206\u5e03\u5728\u6570\u767e\u751a\u81f3\u6570\u5343\u4e2a GPU \u4e0a\u3002\u8fd9\u4e9b GPU \u4f1a\u6301\u7eed\u4ea4\u6362\u6d77\u91cf\u6570\u636e\u2014\u2014\u53c2\u6570\u3001\u68af\u5ea6\u3001\u6fc0\u6d3b\u503c\u2014\u2014\u901a\u5e38\u4ee5\u9ad8\u5ea6\u540c\u6b65\u7684\u7a81\u53d1\u65b9\u5f0f\u8fdb\u884c\u3002\u8fd9\u79cd\u201c\u591a\u5bf9\u4e00\u201d\u6216\u201c\u591a\u5bf9\u4e00\u201d\u7684\u901a\u4fe1\u6a21\u5f0f\uff0c\u88ab\u79f0\u4e3a\u00a0\u88ab\u94f8\u6210<\/strong>\uff0c\u53ef\u4ee5\u8fc5\u901f\u538b\u57ae\u4f20\u7edf\u4ee5\u592a\u7f51\u7f51\u7edc\uff0c\u901a\u5e38\u79f0\u4e3a\u00a0\u541e\u5410\u91cf\u5d29\u6e83<\/strong>\u5982\u679c\u6ca1\u6709\u9002\u5f53\u7684\u673a\u5236\uff0c\u7f51\u7edc\u7f13\u51b2\u533a\u5c31\u4f1a\u6ea2\u51fa\uff0c\u5bfc\u81f4\u6570\u636e\u5305\u4e22\u5931\u3002\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u6570\u636e\u5305\u4e22\u5931\u4e0d\u4ec5\u4ec5\u662f\u5e26\u6765\u4e0d\u4fbf\uff0c\u8fd8\u4f1a\u4e25\u91cd\u964d\u4f4e\u8bad\u7ec3\u6548\u7387\uff0c\u589e\u52a0\u8bad\u7ec3\u65f6\u95f4\uff0c\u751a\u81f3\u5bfc\u81f4\u6a21\u578b\u6536\u655b\u95ee\u9898\u3002\u6570\u636e\u5305\u4e22\u5931\u5bfc\u81f4\u7684\u91cd\u4f20\u4f1a\u9020\u6210\u663e\u8457\u7684\u5ef6\u8fdf\uff0c\u4ece\u800c\u62b5\u6d88 GPU \u7684\u9ad8\u901f\u5904\u7406\u80fd\u529b\u3002<\/p>\n

\u7f51\u7edc\u8fde\u63a5\u3001\u62e5\u585e\u548c GPU \u5229\u7528\u7387<\/h1>\n

\u56e0\u6b64\uff0c\u7f51\u7edc\u7684\u6210\u529f\u6210\u4e3a\u9ad8\u6548\u5229\u7528\u8fd9\u4e9b\u8d44\u672c\u5bc6\u96c6\u578b GPU \u7684\u5173\u952e\uff0c\u65e0\u8bba\u662f\u7528\u4e8e\u8bad\u7ec3\u8fd8\u662f\u63a8\u7406\uff1a<\/p>\n

    \n
  • LLM \u63a8\u7406\uff08\u6a21\u578b\u5e76\u884c\u548c\u6279\u5904\u7406\uff09\uff1a\u00a0<\/strong>\u5bf9\u4e8e\u9700\u8981\u6a21\u578b\u5e76\u884c\u6027\u7684\u5927\u578b LLM\uff08\u5176\u4e2d\u6a21\u578b\u7684\u4e0d\u540c\u5c42\u6216\u90e8\u5206\u4f4d\u4e8e\u4e0d\u540c\u7684 GPU\uff0c\u53ef\u80fd\u4f4d\u4e8e\u4e0d\u540c\u7684\u670d\u52a1\u5668\u4e0a\uff09\uff0c\u6bcf\u4e2a\u63a8\u7406\u8bf7\u6c42\u90fd\u6d89\u53ca\u5728 GPU \u4e4b\u95f4\u8fdb\u884c\u987a\u5e8f\u6570\u636e\u4f20\u8f93\uff0c\u56e0\u4e3a\u63d0\u793a\u4f1a\u6d41\u7ecf\u6a21\u578b\u7684\u5404\u4e2a\u5c42\u3002\u5982\u679c\u8fd9\u4e9b GPU \u4e4b\u95f4\u7684\u7f51\u7edc\u8def\u5f84\u51fa\u73b0\u77ed\u6682\u62e5\u5835\uff0c\u6574\u4e2a\u63a8\u7406\u6d41\u6c34\u7ebf\u5c31\u4f1a\u505c\u6ede\u3002\u540c\u6837\uff0c\u5f53\u4f7f\u7528\u6279\u5904\u7406\u6765\u6700\u5927\u5316 GPU \u5229\u7528\u7387\u65f6\uff0c\u5373\u4f7f\u7531\u4e8e\u6279\u5904\u7406\u4e2d\u67d0\u4e2a\u90e8\u5206\u62e5\u5835\u800c\u5bfc\u81f4\u7684\u5c11\u91cf\u6570\u636e\u5305\u4e22\u5931\u6216\u5ef6\u8fdf\uff0c\u4e5f\u53ef\u80fd\u4f1a\u5ef6\u8fdf\u6574\u4e2a\u6279\u5904\u7406\u7684\u5b8c\u6210\uff0c\u4ece\u800c\u5bf9\u540e\u7eed\u6279\u5904\u7406\u9020\u6210\u8fde\u9501\u53cd\u5e94\u3002\u8fd9\u76f4\u63a5\u5bfc\u81f4\u6700\u7ec8\u7528\u6237\u7684\u63a8\u7406\u5ef6\u8fdf\u589e\u52a0\uff0c\u5e76\u663e\u8457\u964d\u4f4e\u6574\u4e2a\u7cfb\u7edf\u7684\u63a8\u7406\u541e\u5410\u91cf\uff08\u6bcf\u79d2\u67e5\u8be2\u6b21\u6570\uff09\u3002DCB \u7684\u65e0\u635f\u529f\u80fd\u53ef\u786e\u4fdd\u8fd9\u4e9b\u5173\u952e\u7684 GPU \u95f4\u4f20\u8f93\u6c38\u4e0d\u4e2d\u65ad\uff0c\u4ece\u800c\u4fdd\u6301\u6d41\u7545\u3001\u9ad8\u541e\u5410\u91cf\u7684\u63a8\u7406\u6d41\u6c34\u7ebf\uff0c\u5e76\u6700\u5927\u5316\u60a8\u7684 GPU \u6295\u8d44\u56de\u62a5\u7387\u3002<\/li>\n<\/ul>\n
      \n
    • LLM \u5fae\u8c03\uff08\u5206\u5e03\u5f0f\u8bad\u7ec3\uff09\uff1a<\/strong> \u5728\u8de8\u6570\u767e\u4e2a GPU \u7684\u5206\u5e03\u5f0f\u5fae\u8c03\u4f5c\u4e1a\u4e2d\uff0c\u8be5\u8fc7\u7a0b\u6d89\u53ca\u9891\u7e41\u4e14\u5927\u91cf\u7684\u68af\u5ea6\u66f4\u65b0\u548c\u6a21\u578b\u53c2\u6570\u4ea4\u6362\uff08\u4f8b\u5982\uff0cAll-Reduce \u64cd\u4f5c\uff09\u3002\u5982\u679c\u7f51\u7edc\u51fa\u73b0\u62e5\u585e\uff0c\u8fd9\u4e9b\u96c6\u4f53\u901a\u4fe1\u64cd\u4f5c\u7684\u901f\u5ea6\u4f1a\u663e\u8457\u964d\u4f4e\u3002\u4f5c\u4e3a\u5e76\u884c\u5904\u7406\u5668\uff0cGPU \u4f1a\u5904\u4e8e\u7a7a\u95f2\u72b6\u6001\uff0c\u7b49\u5f85\u5176\u4ed6 GPU \u7684\u6570\u636e\u5b8c\u6210\u5f53\u524d\u8fed\u4ee3\u540e\u624d\u80fd\u5f00\u59cb\u4e0b\u4e00\u6b21\u8fed\u4ee3\u3002\u8ba1\u7b97\u5229\u7528\u7387\u663e\u793a\u4e3a 100% \u7684 GPU \u53ef\u80fd\u4f1a\u56e0\u7f51\u7edc I\/O \u800c\u505c\u6ede\uff0c\u8fd9\u610f\u5473\u7740\u5b9e\u9645\u5b8c\u6210\u7684\u5de5\u4f5c\u91cf\u4f1a\u5927\u5927\u51cf\u5c11\uff0c\u4ece\u800c\u6d6a\u8d39\u6570\u5c0f\u65f6\u751a\u81f3\u6570\u5929\u7684\u8ba1\u7b97\u65f6\u95f4\uff0c\u5e76\u589e\u52a0\u4e91\/\u7535\u529b\u6210\u672c\u3002PFC \u786e\u4fdd\u8fd9\u4e9b\u5173\u952e\u7684 All-Reduce \u6570\u636e\u5305\u6c38\u8fdc\u4e0d\u4f1a\u4e22\u5931\uff0c\u4ece\u800c\u9632\u6b62\u707e\u96be\u6027\u7684\u7f51\u7edc\u901f\u5ea6\u4e0b\u964d\uff0c\u800c ECN \u5219\u4f1a\u4e3b\u52a8\u7ba1\u7406\u6d41\u91cf\uff0c\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u8fd9\u4e9b\u7a7a\u95f2\u7b49\u5f85\u3002<\/li>\n<\/ul>\n

      \u6570\u636e\u4e2d\u5fc3\u6865\u63a5\uff1a\u521b\u5efa\u65e0\u635f\u4eba\u5de5\u667a\u80fd\u7f51\u7edc<\/h1>\n

      \u6570\u636e\u4e2d\u5fc3\u6865\u63a5 (DCB) \u662f\u4e00\u7ec4 IEEE \u6807\u51c6 (802.1Qxx)\uff0c\u65e8\u5728\u589e\u5f3a\u6570\u636e\u4e2d\u5fc3\u73af\u5883\u4e2d\u7684\u4ee5\u592a\u7f51\uff0c\u7279\u522b\u662f\u652f\u6301\u5404\u79cd\u6d41\u91cf\u7c7b\u578b\uff08\u5b58\u50a8\u3001\u7ba1\u7406\u548c\u9ad8\u6027\u80fd\u8ba1\u7b97\uff09\u5171\u5b58\u7684\u878d\u5408\u7f51\u7edc\u3002DCB \u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u5173\u952e\u4f5c\u7528\u5728\u4e8e\u5b83\u80fd\u591f\u521b\u5efa\u00a0\u4f4e\u5ef6\u8fdf\u548c\u65e0\u635f\u4ee5\u592a\u7f51\u7ed3\u6784<\/strong>\u786e\u4fdd\u5173\u952e AI \u6d41\u91cf\u4e0d\u4f1a\u4e22\u5931\u6570\u636e\u5305\u3002\u5b9e\u73b0\u8fd9\u4e00\u70b9\u7684 DCB \u7684\u4e24\u4e2a\u5173\u952e\u7ec4\u4ef6\u662f\u57fa\u4e8e\u4f18\u5148\u7ea7\u7684\u6d41\u91cf\u63a7\u5236\uff08\u529f\u7387\u56e0\u6570\u6821\u6b63<\/strong>) \u548c\u663e\u5f0f\u62e5\u585e\u901a\u77e5 (ECN<\/strong>).<\/p>\n

      \u57fa\u4e8e\u4f18\u5148\u7ea7\u7684\u6d41\u91cf\u63a7\u5236 (PFC)\uff1a\u9632\u6b62\u94fe\u8def\u7ea7\u522b\u7684\u6570\u636e\u5305\u4e22\u5931<\/h1>\n

      PFC (IEEE 802.1Qbb) \u662f\u4e00\u79cd\u94fe\u8def\u7ea7\u6d41\u91cf\u63a7\u5236\u673a\u5236\uff0c\u5b83\u6269\u5c55\u4e86\u4f20\u7edf\u7684\u4ee5\u592a\u7f51 PAUSE \u5e27\u3002\u4e0e\u6682\u505c\u94fe\u8def\u4e0a\u6240\u6709\u6d41\u91cf\u7684\u6807\u51c6 PAUSE \u5e27\u4e0d\u540c\uff0cPFC \u5141\u8bb8\u00a0\u6839\u636e\u670d\u52a1\u7c7b\u522b (CoS) \u4f18\u5148\u7ea7\u9009\u62e9\u6027\u6682\u505c\u6d41\u91cf<\/strong>.<\/p>\n

      \u4ee5\u4e0b\u662f PFC \u5728 GPU \u5bc6\u96c6\u73af\u5883\u4e2d\u7684\u5de5\u4f5c\u539f\u7406\u7684\u7b80\u5316\u89e3\u91ca\uff1a<\/p>\n

        \n
      • \u6d41\u91cf\u5206\u7c7b\uff1a<\/strong>\u00a0AI \u8bad\u7ec3\u6d41\u91cf\u901a\u5e38\u5229\u7528\u878d\u5408\u4ee5\u592a\u7f51\u8fdc\u7a0b\u76f4\u63a5\u5185\u5b58\u8bbf\u95ee (RDMA) (RoCEv2)\uff0c\u5e76\u88ab\u5206\u914d\u7279\u5b9a\u7684\u9ad8\u4f18\u5148\u7ea7 CoS\u3002\u8fd9\u786e\u4fdd\u5176\u88ab\u7f51\u7edc\u89c6\u4e3a\u5173\u952e\u6570\u636e\u3002<\/li>\n
      • \u62e5\u585e\u68c0\u6d4b\uff1a<\/strong>\u00a0\u5f53\u4ea4\u6362\u673a\u7684\u7279\u5b9a CoS \u961f\u5217\uff08\u4f8b\u5982\u4e13\u7528\u4e8e RoCEv2 \u6d41\u91cf\u7684\u961f\u5217\uff09\u7684\u51fa\u53e3\u7f13\u51b2\u533a\u8fbe\u5230\u9884\u5b9a\u4e49\u7684\u9608\u503c\u65f6\uff0c\u8868\u793a\u5373\u5c06\u53d1\u751f\u62e5\u585e\u3002<\/li>\n
      • PFC \u6682\u505c\u5e27\uff1a<\/strong>\u00a0\u62e5\u585e\u7684\u4ea4\u6362\u673a\u4f1a\u5411\u4e0a\u6e38\u4f20\u8f93\u8bbe\u5907\uff08\u53e6\u4e00\u53f0\u4ea4\u6362\u673a\u6216 GPU \u7684\u7f51\u7edc\u63a5\u53e3\u5361 - NIC\uff09\u53d1\u9001\u4e00\u4e2a PFC \u6682\u505c\u5e27\u3002\u6b64\u6682\u505c\u5e27\u4e0e\u62e5\u585e\u7684 CoS \u4f18\u5148\u7ea7\u76f8\u5173\u3002<\/li>\n
      • \u9009\u62e9\u6027\u505c\u6b62\uff1a<\/strong>\u00a0\u6536\u5230 PFC \u6682\u505c\u5e27\u540e\uff0c\u4e0a\u6e38\u8bbe\u5907\u4f1a\u6682\u65f6\u505c\u6b62\u4f20\u8f93\u4ec5\u9488\u5bf9\u8be5\u7279\u5b9a CoS \u4f18\u5148\u7ea7\u7684\u6d41\u91cf\u3002\u540c\u4e00\u94fe\u8def\u4e0a\u7684\u5176\u4ed6\u6d41\u91cf\u7c7b\u522b\u4e0d\u53d7\u5f71\u54cd\u3002<\/li>\n
      • \u7f13\u51b2\u533a\u6062\u590d\u548c\u6062\u590d\uff1a<\/strong>\u00a0\u5f53\u62e5\u585e\u7f13\u51b2\u533a\u8017\u5c3d\u5e76\u4e14\u5176\u5360\u7528\u7387\u4f4e\u4e8e\u6062\u590d\u9608\u503c\u65f6\uff0c\u4ea4\u6362\u673a\u4f1a\u53d1\u9001 PFC \u6062\u590d\u5e27\uff0c\u5411\u4e0a\u6e38\u8bbe\u5907\u53d1\u51fa\u4fe1\u53f7\u4ee5\u91cd\u65b0\u542f\u52a8\u8be5\u4f18\u5148\u7ea7\u7684\u4f20\u8f93\u3002<\/li>\n<\/ul>\n

        PFC \u5bf9 GPU \u7684\u597d\u5904\uff1a<\/h1>\n
          \n
        • \u96f6\u6570\u636e\u5305\u4e22\u5931\uff1a<\/strong>\u00a0\u901a\u8fc7\u5728\u7f13\u51b2\u533a\u6ea2\u51fa\u4e4b\u524d\u6682\u505c\u6d41\u91cf\uff0cPFC \u53ef\u4fdd\u8bc1\u5173\u952e AI \u6570\u636e\u7684\u65e0\u635f\u4f20\u8f93\uff0c\u8fd9\u5bf9\u4e8e\u5206\u5e03\u5f0f GPU \u8ba1\u7b97\u7684\u5b8c\u6574\u6027\u548c\u6548\u7387\u81f3\u5173\u91cd\u8981\u3002<\/li>\n
        • \u4ea4\u901a\u9694\u79bb\uff1a<\/strong>\u00a0\u5b83\u53ef\u4ee5\u9632\u6b62\u9ad8\u4f18\u5148\u7ea7 AI \u6d41\u91cf\u7684\u7a81\u53d1\u5f71\u54cd\u540c\u4e00\u94fe\u8def\u4e0a\u5176\u4ed6\u5bf9\u65f6\u95f4\u4e0d\u592a\u654f\u611f\u7684\u6d41\u91cf\u7c7b\u578b\uff0c\u4ece\u800c\u7ef4\u6301\u6574\u4f53\u7f51\u7edc\u7a33\u5b9a\u6027\u3002<\/li>\n
        • \u53ef\u9884\u6d4b\u7684\u6027\u80fd\uff1a<\/strong>\u00a0\u901a\u8fc7\u6d88\u9664\u6570\u636e\u5305\u4e22\u5931\uff0cPFC \u6709\u52a9\u4e8e\u5b9e\u73b0 GPU \u901a\u4fe1\u66f4\u53ef\u9884\u6d4b\u548c\u4e00\u81f4\u7684\u6027\u80fd\uff0c\u51cf\u5c11\u6296\u52a8\u5e76\u7f29\u77ed\u4f5c\u4e1a\u5b8c\u6210\u65f6\u95f4\u3002<\/li>\n<\/ul>\n

          \u7136\u800c\uff0cPFC \u4e5f\u5b58\u5728\u5c40\u9650\u6027\u3002\u5982\u679c\u8bbe\u8ba1\u548c\u914d\u7f6e\u4e0d\u5f53\uff0c\u53ef\u80fd\u4f1a\u51fa\u73b0\u201cPFC \u98ce\u66b4\u201d\uff0c\u5bfc\u81f4\u6682\u505c\u5e27\u5e7f\u6cdb\u4f20\u64ad\uff0c\u6700\u7ec8\u53ef\u80fd\u5bfc\u81f4\u6574\u4e2a\u7f51\u7edc\u901f\u5ea6\u53d8\u6162\u6216\u6b7b\u9501\uff0c\u5c24\u5176\u662f\u5728\u591a\u8df3\u73af\u5883\u4e2d\u3002\u00a0\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u8981\u6dfb\u52a0\u53e6\u4e00\u9879\u6280\u672f ECN \u6765\u8865\u5145 PFC<\/strong>.<\/p>\n

          \u663e\u5f0f\u62e5\u585e\u901a\u77e5\uff08ECN\uff09\uff1a\u4e3b\u52a8\u62e5\u585e\u907f\u514d<\/h1>\n

          ECN\uff08RFC 3168\uff09\u662f\u4e00\u79cd\u5141\u8bb8\u7f51\u7edc\u8bbe\u5907\u5411\u7aef\u70b9\u53d1\u51fa\u65e9\u671f\u62e5\u585e\u4fe1\u53f7\u7684\u673a\u5236\u00a0\u524d<\/strong>\u00a0\u53d1\u751f\u6570\u636e\u5305\u4e22\u5931\u65f6\uff0c\u652f\u6301 ECN \u7684\u8bbe\u5907\u4e0d\u4f1a\u4e22\u5f03\u6570\u636e\u5305\uff0c\u800c\u662f\u5728 IP \u62a5\u5934\u4e2d\u6807\u8bb0\u6570\u636e\u5305\u4ee5\u6307\u793a\u62e5\u585e\u3002\u5b83\u63d0\u4f9b\u4e86\u53ef\u9760\u7684 GPU \u5230 GPU \u6d41\u91cf\u6240\u9700\u7684\u4e3b\u52a8\u961f\u5217\u7ba1\u7406\u7ea7\u522b\u3002<\/p>\n

          ECN\u6d41\u91cf\u7ba1\u7406\u6d41\u7a0b\u901a\u5e38\u5982\u4e0b\uff1a<\/p>\n

            \n
          • ECN \u534f\u5546\u80fd\u529b\uff1a<\/strong>\u00a0\u5728\u8fde\u63a5\u5efa\u7acb\u671f\u95f4\uff08\u4f8b\u5982\uff0cTCP \u63e1\u624b\uff09\uff0c\u53d1\u9001\u65b9\u548c\u63a5\u6536\u65b9\u534f\u5546\u5176 ECN \u529f\u80fd\u3002<\/li>\n
          • \u62e5\u5835\u6807\u8bb0\uff1a<\/strong>\u00a0\u5f53\u7f51\u7edc\u8bbe\u5907\u7684\u961f\u5217\u5229\u7528\u7387\u8fbe\u5230 ECN \u9608\u503c\uff08\u4f4e\u4e8e\u89e6\u53d1 PFC \u7684\u9608\u503c\uff09\u65f6\uff0c\u8bbe\u5907\u4f1a\u5c06\u4f20\u5165\u7684\u652f\u6301 ECN \u7684\u6570\u636e\u5305\u6807\u8bb0\u4e3a\u201c\u9047\u5230\u62e5\u585e\u201d\uff08CE\uff09\u3002<\/li>\n
          • \u63a5\u6536\u8005\u901a\u77e5\uff1a<\/strong>\u00a0\u6807\u8bb0\u7684\u6570\u636e\u5305\u5230\u8fbe\u5177\u6709 ECN \u529f\u80fd\u7684\u63a5\u6536\u65b9\u3002<\/li>\n
          • \u53d1\u4ef6\u4eba\u53cd\u9988\uff1a<\/strong>\u00a0\u7136\u540e\uff0c\u63a5\u6536\u65b9\u5c06\u6b64\u62e5\u585e\u901a\u77e5\u56de\u663e\u7ed9\u53d1\u9001\u65b9\uff08\u4f8b\u5982\uff0c\u901a\u8fc7\u5728 TCP \u6807\u5934\u4e2d\u8bbe\u7f6e ECN-Echo (ECE) \u4f4d\u6216\u5728 RoCEv2 \u4e2d\u53d1\u9001\u62e5\u585e\u901a\u77e5\u6570\u636e\u5305 (CNP)\uff09\u3002<\/li>\n
          • \u964d\u4f4e\u5229\u7387\uff1a<\/strong>\u00a0\u53d1\u9001\u65b9\u6536\u5230\u62e5\u585e\u53cd\u9988\u540e\uff0c\u4e3b\u52a8\u964d\u4f4e\u53d1\u9001\u901f\u7387\uff0c\u4ece\u800c\u5728\u7f13\u51b2\u533a\u6ea2\u51fa\u548c\u4e22\u5305\u4e4b\u524d\u7f13\u89e3\u62e5\u585e\u3002<\/li>\n<\/ul>\n

            PFC \u548c ECN \u534f\u540c\u4f5c\u7528\uff0c\u9002\u7528\u4e8e\u5927\u578b GPU \u90e8\u7f72\uff1a<\/h1>\n

            \u5728\u7531\u6570\u767e\u6216\u6570\u5343\u4e2a\u6602\u8d35\u7684 GPU \u7ec4\u6210\u7684\u5927\u89c4\u6a21 AI \u96c6\u7fa4\u4e2d\uff0cPFC \u548c ECN \u534f\u540c\u5de5\u4f5c\uff0c\u63d0\u4f9b\u5f3a\u5927\u800c\u9ad8\u6548\u7684\u65e0\u635f\u7f51\u7edc\uff0c\u4ece\u800c\u63d0\u9ad8 GPU \u672c\u8eab\u7684\u4ea4\u4ed8\u4ef7\u503c\uff1a<\/p>\n

              \n
            • ECN \u4f5c\u4e3a\u7b2c\u4e00\u9053\u9632\u7ebf\uff1a<\/strong>\u00a0ECN \u662f\u4e00\u79cd\u4e3b\u52a8\u673a\u5236\uff0c\u80fd\u591f\u63d0\u4f9b\u62e5\u585e\u9884\u8b66\u3002\u901a\u8fc7\u5141\u8bb8\u53d1\u9001\u65b9\u9884\u5148\u964d\u4f4e\u901f\u7387\uff0c\u5b83\u53ef\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u964d\u4f4e\u8fbe\u5230 PFC \u9608\u503c\u7684\u53ef\u80fd\u6027\uff0c\u5e76\u907f\u514d\u91c7\u53d6\u66f4\u4e3a\u6fc0\u8fdb\u7684\u6682\u505c\u6d41\u91cf\u63aa\u65bd\u3002\u8fd9\u79cd\u201c\u8f6f\u201d\u901f\u7387\u81ea\u9002\u5e94\u5bf9\u4e8e\u7ef4\u6301\u8fde\u7eed\u7684\u6570\u636e\u6d41\u81f3\u5173\u91cd\u8981\u3002<\/li>\n
            • PFC \u4f5c\u4e3a\u6700\u540e\u624b\u6bb5\uff1a<\/strong>\u00a0\u5982\u679c ECN \u7684\u4e3b\u52a8\u63aa\u65bd\u4e0d\u8db3\u4ee5\u9632\u6b62\u62e5\u585e\uff0c\u6216\u8005\u7a81\u7136\u53d1\u751f\u5927\u91cf\u6d41\u91cf\uff0cPFC \u5c06\u4f5c\u4e3a\u4e00\u79cd\u88ab\u52a8\u7684\u786c\u505c\u6b62\u673a\u5236\u4ecb\u5165\uff0c\u4ee5\u9632\u6b62\u6700\u5173\u952e\u7684 AI \u6d41\u91cf\u53d1\u751f\u4efb\u4f55\u6570\u636e\u5305\u4e22\u5931\u3002<\/li>\n
            • \u4f18\u5316 RoCEv2 \u6027\u80fd\uff1a<\/strong>\u00a0RoCEv2 \u88ab\u5e7f\u6cdb\u7528\u4e8e GPU \u4e92\u8fde\uff0c\u5b83\u9ad8\u5ea6\u4f9d\u8d56\u4e8e\u8fd9\u4e9b\u673a\u5236\u3002ECN \u4fe1\u53f7\u4f1a\u89e6\u53d1\u7f51\u5361 (NIC) \u5185\u7684\u62e5\u585e\u63a7\u5236\u7b97\u6cd5\uff08\u4f8b\u5982\u6570\u636e\u4e2d\u5fc3\u91cf\u5316\u62e5\u585e\u901a\u77e5 - DCQCN\uff09\uff0c\u4ece\u800c\u52a8\u6001\u8c03\u6574\u4f20\u8f93\u901f\u7387\u3002PFC \u786e\u4fdd\u5373\u4f7f\u5728\u6781\u7aef\u8d1f\u8f7d\u4e0b\uff0cRoCEv2 \u6570\u636e\u5305\u4e5f\u4e0d\u4f1a\u4e22\u5931\uff0c\u4ece\u800c\u7ef4\u62a4 RDMA \u64cd\u4f5c\u7684\u5b8c\u6574\u6027\u3002<\/li>\n
            • \u5e73\u8861\u5ef6\u8fdf\u548c\u541e\u5410\u91cf\uff1a<\/strong> \u901a\u8fc7\u5c06 ECN \u7684\u4e3b\u52a8\u901f\u7387\u9650\u5236\u4e0e PFC \u7684\u65e0\u635f\u4fdd\u8bc1\u76f8\u7ed3\u5408\uff0c\u7f51\u7edc\u67b6\u6784\u5e08\u53ef\u4ee5\u5bf9\u7f51\u7edc\u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u5e73\u8861\u5bf9\u5ef6\u8fdf\u654f\u611f\u7684\u201c\u5c0f\u9f20\u6d41\u201d\uff08\u5c0f\u578b\u4ea4\u4e92\u5f0f\u6d88\u606f\uff09\u7684\u4f4e\u5ef6\u8fdf\u548c\u5927\u578b\u201c\u5927\u8c61\u6d41\u201d\uff08\u8bad\u7ec3\u671f\u95f4\u7684\u6279\u91cf\u6570\u636e\u4f20\u8f93\uff09\u7684\u9ad8\u541e\u5410\u91cf\u3002<\/li>\n<\/ul>\n

              \u771f\u5b9e\u4f01\u4e1a\u793a\u4f8b\uff1a\u5177\u6709\u4e0d\u540c QoS \u548c SLA \u8981\u6c42\u7684\u591a\u79df\u6237 AI \u5de5\u4f5c\u8d1f\u8f7d<\/h1>\n

              \u5728\u5927\u578b\u4f01\u4e1a\u4e2d\uff0cAI \u96c6\u7fa4\u5f88\u5c11\u4e13\u7528\u4e8e\u5355\u4e00\u4efb\u52a1\u3002\u8bbe\u60f3\u4e00\u4e2a\u591a\u79df\u6237\u73af\u5883\uff0c\u5176\u4e2d\u540c\u4e00\u4e2a GPU \u96c6\u7fa4\u652f\u6301\u591a\u4e2a\u4e0d\u540c\u7684 AI \u5de5\u4f5c\u8d1f\u8f7d\uff1a<\/p>\n

                \n
              • \u7814\u53d1\uff08R&D\uff09\u56e2\u961f\uff1a<\/strong> \u9488\u5bf9\u65b0\u4ea7\u54c1\u529f\u80fd\u8fd0\u884c\u5b9e\u9a8c\u6027\u7684 LLM \u5fae\u8c03\u4f5c\u4e1a\u3002\u8fd9\u4e9b\u4f5c\u4e1a\u901a\u5e38\u89c4\u6a21\u5e9e\u5927\u3001\u8fd0\u884c\u65f6\u95f4\u957f\uff0c\u53ef\u4ee5\u5bb9\u5fcd\u7a0d\u9ad8\u7684\u521d\u59cb\u5ef6\u8fdf\uff0c\u4f46\u9700\u8981\u4fdd\u8bc1\u5e26\u5bbd\u624d\u80fd\u5728\u7279\u5b9a\u65f6\u95f4\u7a97\u53e3\u5185\u5b8c\u6210\uff08\u4f8b\u5982\uff0c\u591c\u95f4\uff0c\u5e76\u6839\u636e SLA \u8981\u6c42\u5728\u6b21\u65e5\u65e9\u6668\u524d\u5b8c\u6210\uff09\u3002\u5bf9\u4e8e\u8fd9\u4e9b\u4f5c\u4e1a\u800c\u8a00\uff0c\u7a33\u5b9a\u7684\u541e\u5410\u91cf\u4ee5\u907f\u514d\u8bad\u7ec3\u5ef6\u8fdf\u81f3\u5173\u91cd\u8981\uff0c\u5373\u4f7f\u8fd9\u610f\u5473\u7740\u4f1a\u6682\u65f6\u964d\u4f4e\u5176\u4ed6\u4f4e\u4f18\u5148\u7ea7\u6d41\u91cf\u7684\u901f\u5ea6\u3002<\/li>\n
              • \u5236\u9020\u8d28\u91cf\u63a7\u5236\uff08QC\uff09\u90e8\u95e8\uff1a <\/strong>\u9700\u8981\u5728\u88c5\u914d\u7ebf\u4e0a\u8fdb\u884c\u5b9e\u65f6\u7f3a\u9677\u68c0\u6d4b\uff0c\u8981\u6c42\u6781\u4f4e\u7684\u5ef6\u8fdf\uff08<10ms\uff09\u548c\u63a5\u8fd1\u96f6\u7684\u6570\u636e\u5305\u4e22\u5931\uff08\u4f8b\u5982\uff0c99.999% \u6210\u529f\u63a8\u7406\u7684 SLA\uff09\uff0c\u4ee5\u9632\u6b62\u751f\u4ea7\u7ebf\u505c\u5de5\u3002<\/li>\n
              • \u8d22\u52a1\u98ce\u9669\u5206\u6790\u56e2\u961f\uff1a <\/strong>\u8fd0\u884c\u6279\u91cf\u63a8\u7406\u4f5c\u4e1a\uff0c\u7528\u4e8e\u6b3a\u8bc8\u68c0\u6d4b\u6216\u5e02\u573a\u9884\u6d4b\u3002\u8fd9\u4e9b\u4efb\u52a1\u81f3\u5173\u91cd\u8981\uff0c\u4f46\u4ea4\u4e92\u6027\u8f83\u5dee\uff0c\u9700\u8981\u9ad8\u541e\u5410\u91cf\u624d\u80fd\u5728\u7279\u5b9a\u7684\u5408\u89c4\u65f6\u95f4\u6bb5\u5185\uff08\u4f8b\u5982\uff0c\u4e00\u5929\u7ed3\u675f\u524d\uff09\u5904\u7406\u5927\u578b\u6570\u636e\u96c6\u3002<\/li>\n<\/ul>\n

                DCB \u7ed3\u5408 PFC \u548c ECN\uff0c\u5141\u8bb8\u7f51\u7edc\u7ba1\u7406\u5458\u5bf9\u8fd9\u4e9b\u4e0d\u540c\u7684\u6d41\u91cf\u7c7b\u578b\u8fdb\u884c\u5206\u7c7b\u548c\u4f18\u5148\u7ea7\u6392\u5e8f\u3002\u5bf9\u4e8e\u5ef6\u8fdf\u654f\u611f\u7684\u63a8\u7406\u67e5\u8be2\uff0cRoCEv2 \u6d41\u91cf\u53ef\u4ee5\u5206\u914d\u6700\u9ad8\u4f18\u5148\u7ea7\u7684 CoS\uff0c\u5e76\u8bbe\u7f6e\u4e25\u683c\u7684 ECN \u9608\u503c\uff0c\u4ee5\u786e\u4fdd\u4e3b\u52a8\u964d\u4f4e\u901f\u7387\uff1b\u800c\u8bad\u7ec3\u6d41\u91cf\u5219\u83b7\u5f97\u9ad8\u4f18\u5148\u7ea7\u7684 CoS\uff0c\u4ee5\u5b9e\u73b0\u65e0\u635f\u4f20\u8f93\uff1b\u5176\u4ed6\u540e\u53f0\u6d41\u91cf\u5219\u4f7f\u7528\u6807\u51c6\u7c7b\u522b\u3002\u8fd9\u79cd\u5148\u8fdb\u7684\u6d41\u91cf\u7ba1\u7406\u673a\u5236\u53ef\u786e\u4fdd\u9ad8\u4f18\u5148\u7ea7\u7684\u63a8\u7406\u67e5\u8be2\u4e0d\u4f1a\u56e0\u5927\u89c4\u6a21\u8bad\u7ec3\u4f5c\u4e1a\u800c\u5904\u4e8e\u8d44\u6e90\u532e\u4e4f\u7684\u72b6\u6001\uff0c\u4ece\u800c\u786e\u4fdd\u5728\u5171\u4eab GPU \u57fa\u7840\u67b6\u6784\u4e0a\u6ee1\u8db3\u6bcf\u4e2a\u79df\u6237\u7684 SLA \u8981\u6c42\u3002<\/p>\n

                \u7ed3\u8bba<\/h1>\n

                \u968f\u7740\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u7684\u590d\u6742\u6027\u548c\u89c4\u6a21\u4e0d\u65ad\u589e\u957f\uff0c\u5e95\u5c42\u7684\u6a2a\u5411\u6269\u5c55\u7f51\u7edc\u57fa\u7840\u8bbe\u65bd\u5df2\u6210\u4e3a\u6574\u4f53\u7cfb\u7edf\u6027\u80fd\u4e2d\u65e5\u76ca\u5173\u952e\u7684\u7ec4\u6210\u90e8\u5206\u3002\u6570\u636e\u4e2d\u5fc3\u6865\u63a5\u6280\u672f\u4ee5\u5176\u57fa\u4e8e\u4f18\u5148\u7ea7\u7684\u6d41\u91cf\u63a7\u5236 (PFC) \u548c\u663e\u5f0f\u62e5\u585e\u901a\u77e5 (ECN) \u7b49\u6838\u5fc3\u529f\u80fd\u4e3a\u57fa\u77f3\uff0c\u5b83\u4e0d\u4ec5\u4ec5\u662f\u4e00\u9879\u4f18\u5316\u6216\u9526\u4e0a\u6dfb\u82b1\u7684\u529f\u80fd\uff1b\u5b83\u66f4\u662f\u6784\u5efa\u9ad8\u6548\u3001\u53ef\u9760\u4e14\u53ef\u6269\u5c55\u7684\u4eba\u5de5\u667a\u80fd\u8ba1\u7b97\u89e3\u51b3\u65b9\u6848\u7684\u5173\u952e\u57fa\u7840\uff0c\u5c24\u5176\u662f\u5728 GPU \u65b9\u9762\u6295\u5165\u5de8\u5927\u7684\u60c5\u51b5\u4e0b\u3002\u901a\u8fc7\u9009\u62e9\u80fd\u591f\u786e\u4fdd\u65e0\u635f\u901a\u4fe1\u3001\u4e3b\u52a8\u7ba1\u7406\u62e5\u585e\u5e76\u63d0\u4f9b\u7cbe\u7ec6\u6d41\u91cf\u63a7\u5236\u7684\u7f51\u7edc\u8fde\u63a5\u89e3\u51b3\u65b9\u6848\uff0c\u57fa\u4e8e\u8fd9\u4e9b\u6570\u636e\u4e2d\u5fc3\u6865\u63a5\u6280\u672f\u7684\u53ef\u6269\u5c55\u89e3\u51b3\u65b9\u6848\u5c06\u8d4b\u80fd\u4f01\u4e1a\u53ca\u5176\u4eba\u5de5\u667a\u80fd\u6218\u7565\uff0c\u4f7f\u5176\u80fd\u591f\u5229\u7528\u884c\u4e1a\u6807\u51c6\u6280\u672f\u7a81\u7834\u6781\u9650\uff0c\u52a0\u901f\u4eba\u5de5\u667a\u80fd\u65f6\u4ee3\u7684\u521b\u65b0\u6b65\u4f10\u3002<\/p>\n<\/div><\/div><\/div><\/div><\/div>

                \u5982\u679c\u60a8\u5bf9\u6211\u4eec\u7684\u4ea7\u54c1\u548c\u670d\u52a1\u6709\u4efb\u4f55\u610f\u89c1\u3001\u8be2\u95ee\u6216\u95ee\u9898\uff0c\u8bf7\u586b\u5199\u4ee5\u4e0b\u8868\u683c\u3002<\/p>\n<\/div>