{"id":34716,"date":"2024-12-11T14:57:39","date_gmt":"2024-12-11T06:57:39","guid":{"rendered":"https:\/\/17aitech.com\/?p=34716"},"modified":"2024-12-11T14:57:39","modified_gmt":"2024-12-11T06:57:39","slug":"%e4%bb%8e%e8%a3%b8%e6%9c%ba%e5%88%b0700%e4%ba%bf%e5%8f%82%e6%95%b0%e5%a4%a7%e6%a8%a1%e5%9e%8b%ef%bc%8c%e8%bf%99%e9%87%8c%e6%9c%89%e4%bb%bd%e6%95%99%e7%a8%8b%ef%bc%8c%e8%bf%98%e6%9c%89%e7%8e%b0","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34716","title":{"rendered":"\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-24-4\" target=\"_blank\">\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c<\/a><\/p>\n<p>\u6211\u4eec\u77e5\u9053 LLM \u662f\u5728\u5927\u89c4\u6a21\u8ba1\u7b97\u673a\u96c6\u7fa4\u4e0a\u4f7f\u7528\u6d77\u91cf\u6570\u636e\u8bad\u7ec3\u5f97\u5230\u7684\uff0c<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>\u66fe\u4ecb\u7ecd\u8fc7\u4e0d\u5c11\u7528\u4e8e\u8f85\u52a9\u548c\u6539\u8fdb LLM \u8bad\u7ec3\u6d41\u7a0b\u7684\u65b9\u6cd5\u548c\u6280\u672f\u3002\u800c\u4eca\u5929\uff0c\u6211\u4eec\u8981\u5206\u4eab\u7684\u662f\u4e00\u7bc7\u6df1\u5165\u6280\u672f\u5e95\u5c42\u7684\u6587\u7ae0\uff0c\u4ecb\u7ecd\u5982\u4f55\u5c06\u4e00\u5806\u8fde<mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark>\u4e5f\u6ca1\u6709\u7684\u300c\u88f8\u673a\u300d\u53d8\u6210\u7528\u4e8e\u8bad\u7ec3 LLM \u7684\u8ba1\u7b97\u673a\u96c6\u7fa4\u3002<\/p>\n<p>\u8fd9\u7bc7\u6587\u7ae0\u6765\u81ea\u4e8e AI \u521d\u521b\u516c\u53f8 Imbue\uff0c\u8be5\u516c\u53f8\u81f4\u529b\u4e8e\u901a\u8fc7\u7406\u89e3\u673a\u5668\u7684\u601d\u7ef4\u65b9\u5f0f\u6765\u5b9e\u73b0\u901a\u7528\u667a\u80fd\u3002<\/p>\n<p>\u5f53\u7136\uff0c\u5c06\u4e00\u5806\u8fde<mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark>\u4e5f\u6ca1\u6709\u7684\u300c\u88f8\u673a\u300d\u53d8\u6210\u7528\u4e8e\u8bad\u7ec3 LLM \u7684\u8ba1\u7b97\u673a\u96c6\u7fa4\u5e76\u4e0d\u662f\u4e00\u4e2a\u8f7b\u677e\u7684\u8fc7\u7a0b\uff0c\u5145\u6ee1\u4e86\u63a2\u7d22\u548c\u8bd5\u9519\uff0c\u4f46 Imbue \u6700\u7ec8\u6210\u529f\u8bad\u7ec3\u4e86\u4e00\u4e2a 700 \u4ebf<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7684 LLM\uff0c\u5e76\u5728\u6b64\u8fc7\u7a0b\u4e2d\u79ef\u7d2f\u4e86\u8bb8\u591a\u6709\u7528\u7684\u7ecf\u9a8c\u3002<\/p>\n<p>\u672c\u6587\u5c06\u6df1\u5165\u4ecb\u7ecd\u8be5\u56e2\u961f\u6784\u5efa\u81ea\u5df1\u7684 LLM \u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\u7684\u5168\u8fc7\u7a0b\uff0c\u5e76\u4f1a\u5206\u4eab\u4ed6\u4eec\u4e3a\u65b9\u4fbf\u76d1\u63a7\u3001\u68c0\u67e5\u548c\u7ea0\u9519\u800c\u7f16\u5199\u7684\u8bf8\u591a\u5de5\u5177\u548c\u811a\u672c\u3002<\/p>\n<p>\u5982\u679c\u4f60\u6709\u5fc3\u6784\u5efa\u81ea\u5df1\u7684 LLM \u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\u6216\u597d\u5947 LLM \u662f\u5982\u4f55\u70bc\u6210\u7684\uff0c\u90a3\u4e48\u8fd9\u7bc7\u6587\u7ae0\u503c\u5f97\u4f60\u9605\u8bfb\u548c\u6536\u85cf\u3002<\/p>\n<p>\u4ee5\u4e0b\u662f Imbue \u56e2\u961f\u6587\u7ae0\u539f\u6587\u3002<\/p>\n<p><strong>\u5f15\u8a00<\/strong><\/p>\n<p>\u6211\u4eec\u8fd9\u4e2a\u7531\u7814\u7a76\u8005\u548c\u5de5\u7a0b\u5e08\u7ec4\u6210\u7684\u5c0f\u56e2\u961f\u7528\u4e86\u51e0\u4e2a\u6708\u65f6\u95f4\u5728\u81ea\u5df1\u7684\u57fa\u7840\u8bbe\u65bd\u4e0a\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u4e86\u4e00\u4e2a 700 \u4ebf<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u91cf\u7684\u6a21\u578b\uff0c\u5e76\u4e14\u8be5\u6a21\u578b\u5728\u63a8\u7406\u76f8\u5173\u7684\u4efb\u52a1\u4e0a\u80dc\u8fc7\u4e86\u96f6\u6837\u672c\u7684 GPT-4o\u3002<\/p>\n<p>\u4eca\u5929\uff0c\u6211\u4eec\u8981\u5206\u4eab\u7684\u662f\u8bbe\u7f6e\u6240\u9700\u57fa\u7840\u8bbe\u65bd\u7684\u8fc7\u7a0b\uff1a\u4ece\u7ec4\u5408\u521d\u59cb\u96c6\u7fa4\u548c\u5b89\u88c5<mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark>\u5230\u8bbe\u7f6e\u5728\u8bad\u7ec3\u671f\u95f4\u906d\u9047\u9519\u8bef\u65f6\u81ea\u52a8\u6062\u590d\u3002\u6211\u4eec\u4f1a\u8be6\u7ec6\u8bf4\u660e\u5728\u6bcf\u4e00\u6b65\u906d\u9047\u5230\u7684\u96be\u9898\u548c\u89e3\u51b3\u65b9\u6cd5\u3002\u9664\u4e86\u8fd9\u4e9b\u5fc3\u5f97\u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u5c06\u53d1\u5e03\u6211\u4eec\u4e00\u8def\u4e0a\u5f00\u53d1\u7684\u8bb8\u591a\u811a\u672c\uff0c\u4ee5\u4fbf\u5176\u4ed6\u56e2\u961f\u80fd\u66f4\u8f7b\u677e\u5730\u4e3a\u81ea\u5df1\u7684\u6a21\u578b\u8bad\u7ec3\u521b\u5efa\u7a33\u5b9a\u7684\u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n<p>\u5728\u6574\u4e2a\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u7684\u5de5\u7a0b\u5e08\u56e2\u961f\u4e0e Voltage Park \u4e00\u8d77\u51c6\u5907\u597d\u4e86\u8ba1\u7b97\u673a\u96c6\u7fa4\uff0c\u6784\u5efa\u4e86\u751f\u4ea7\u5e94\u7528\u7684\u57fa\u7840\u3002\u8fd9\u6574\u4e2a\u8fc7\u7a0b\u5305\u62ec\uff1a<\/p>\n<p>1. \u914d\u7f6e\u5404\u53f0\u673a\u5668<\/p>\n<p>2. \u914d\u7f6e\u00a0InfiniBand<\/p>\n<p>3. \u786e\u4fdd\u673a\u5668\u5b8c\u5168\u5065\u5eb7<\/p>\n<p>4. \u8bca\u65ad\u5e38\u89c1\u7684\u8bad\u7ec3\u95ee\u9898<\/p>\n<p>5. \u6539\u8fdb\u57fa\u7840\u8bbe\u65bd\u5de5\u5177<\/p>\n<p>\u4e0b\u9762\u5c06\u8be6\u7ec6\u63cf\u8ff0\u6bcf\u4e2a\u6b65\u9aa4\u3002<\/p>\n<p><strong>\u80cc\u666f\uff1a\u8fd9\u662f\u5982\u4f55\u70bc\u6210\u7684<\/strong><\/p>\n<p>\u6211\u4eec\u6267\u884c\u8ba1\u7b97\u7684\u76ee\u6807\u662f\u786e\u4fdd\u80fd\u5feb\u901f\u5b9e\u9a8c\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u9700\u8981\u5927\u91cf\u9ad8\u901f GPU\uff0c\u5e76\u4e14\u8fd9\u4e9b GPU \u4e4b\u95f4\u4e5f\u80fd\u9ad8\u901f\u901a\u4fe1\u3002<\/p>\n<p>\u672c\u6587\u5c06\u91cd\u70b9\u5173\u6ce8\u4e00\u4e2a\u96c6\u7fa4\uff0c\u5176\u4e2d\u5305\u542b\u5206\u6563\u5728 511 \u53f0\u8ba1\u7b97\u673a\u4e2d\u7684 4088 \u53f0 H100 GPU\uff0c\u4e5f\u5c31\u662f\u6bcf\u53f0\u8ba1\u7b97\u673a 8 \u53f0 GPU\u3002\u4e4b\u6240\u4ee5\u6709 511 \u53f0\u5e26 GPU \u7684\u8ba1\u7b97\u673a\uff0c\u662f\u56e0\u4e3a\u9700\u8981\u4fdd\u7559\u4e00\u4e9b\u8fde\u63a5\u7ed9\u7edf\u4e00\u7ed3\u6784\u7ba1\u7406\u5668\uff08Unified Fabric Manager\uff09\u8282\u70b9\uff0c\u5176\u4f5c\u7528\u662f\u7ba1\u7406 InfiniBand \u7f51\u7edc\u3002\u5728 511 \u53f0\u5e26 GPU \u7684\u4e3b\u673a\u4e0a\uff0c\u6bcf\u53f0 GPU \u90fd\u76f4\u63a5\u4e0e\u4e00\u5757 ConnectX-7 \u7f51\u5361\u76f8\u8fde\uff0c\u5176\u80fd\u4ee5 400 Gbps \u7684\u901f\u5ea6\u4e0e\u8be5 InfiniBand \u7f51\u7edc\u4e2d\u7684\u4efb\u610f GPU \u4f20\u8f93\u6570\u636e\u3002<\/p>\n<p>\u6211\u4eec\u7684 InfiniBand \u7f51\u7edc\u62d3\u6251\u7ed3\u6784\u7684\u5f62\u5f0f\u662f\u300cfully non-blocking\u300d\uff0c\u5373\u5b8c\u5168\u65e0\u963b\u585e\uff1b\u7406\u8bba\u4e0a\u8bb2\uff0c\u8fd9\u80fd\u8ba9 GPU \u4ee5\u6700\u5927\u901f\u5ea6\u4e92\u76f8\u901a\u4fe1\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u4f7f\u7528\u4e86\u4e00\u79cd\u4e09\u5c42\u5f0f InfiniBand \u7f51\u7edc\u67b6\u6784\uff1a\u4e09\u5c42 InfiniBand \u4ea4\u6362\u673a\u3002\u53ea\u8981\u8fde\u63a5\u6b63\u786e\uff0c\u4fbf\u80fd\u8ba9\u6574\u4e2a\u7f51\u7edc\u90fd\u83b7\u5f97\u8fd9\u6837\u9ad8\u6c34\u5e73\u7684\u541e\u5410\u91cf\u3002\u4e0b\u56fe\u5c55\u793a\u4e86\u8fd9\u4e2a InfiniBand \u7f51\u7edc\u7684\u6982\u51b5\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ff6478d07102151bce7de146029da7a7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ff6478d07102151bce7de146029da7a7.png\"><\/a><\/p>\n<p>\u8bf7\u6ce8\u610f\uff0c\u8bad\u7ec3\u7f51\u7edc\u65f6\u7684\u901a\u4fe1\u53d1\u751f\u5728 InfiniBand \u4e0a\uff0c\u800c\u4e0d\u662f\u4ee5\u592a\u7f51\u4e0a\u3002\u5c3d\u7ba1\u8fd9\u4e9b\u673a\u5668\u4e5f\u8fde\u63a5\u4e86\u4ee5\u592a\u7f51\uff0c\u4f46\u8be5\u7f51\u7edc\u7684\u4f5c\u7528\u662f\u4f20\u8f93\u6570\u636e\u96c6\u548c\u68c0\u67e5\u70b9\u7b49\u6570\u636e\u3002\u5982\u679c\u4f7f\u7528\u4ee5\u592a\u7f51\u6765\u53d1\u9001\u6570\u636e\uff0c\u901f\u5ea6\u4f1a\u6162\u5f97\u591a\uff0c\u56e0\u4e3a\u6570\u636e\u4f1a\u5148\u4ece GPU \u4f20\u8f93\u5230 CPU\uff0c\u7136\u540e\u518d\u901a\u8fc7 100 Gbps \u901f\u5ea6\u7684\u4ee5\u592a\u7f51\u5361\u53d1\u51fa\u53bb\u3002\u5c3d\u7ba1\u4e5f\u53ef\u4ee5\u4f7f\u7528\u540d\u4e3a RDMA over Converged Ethernet\uff08RoCE\uff09\u7684\u6280\u672f\u57fa\u4e8e\u4ee5\u592a\u7f51\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f46\u90a3\u9700\u8981\u5728\u786c\u4ef6\u548c\u8f6f\u4ef6\u65b9\u9762\u90fd\u505a\u5927\u91cf\u989d\u5916\u5de5\u4f5c\uff0c\u5e76\u4e14\u53ef\u9760\u7a0b\u5ea6\u901a\u5e38\u4e5f\u4e0d\u53ca InfiniBand\u3002\u8be6\u60c5\u53ef\u53c2\u9605\u8fd9\u7bc7\u8bba\u6587\uff1ahttps:\/\/arxiv.org\/pdf\/2402.15627<\/p>\n<p>\u53e6\u5916\u8fd8\u6709\u4e00\u4e2a\u4ec5\u7528\u4e8e\u914d\u7f6e\u548c\u7ba1\u7406\u7684\u8f85\u52a9\u4ee5\u592a\u7f51\uff0c\u4ece\u800c\u53ef\u8bbf\u95ee BIOS\uff08\u57fa\u672c\u8f93\u5165\u8f93\u51fa\u7cfb\u7edf\uff09\u3001\u7535\u6e90\u548c\u5176\u4ed6\u4f4e\u5c42\u673a\u5668\u63a5\u53e3\u7684\u63a7\u5236\u754c\u9762\u3002\u5982\u679c\u6ca1\u6709\u8fd9\u4e2a\u7ba1\u7406\u7f51\u7edc\uff0c\u6211\u4eec\u5c31\u5fc5\u987b\u901a\u8fc7 USB \u9a71\u52a8\u3001\u952e\u76d8\u548c\u663e\u793a\u5668\u6765\u624b\u52a8\u8bbe\u7f6e\u6bcf\u4e2a\u8282\u70b9\u3002\u5bf9\u4e8e\u6709\u51e0\u767e\u53f0\u673a\u5668\u7684\u60c5\u51b5\uff0c\u8fd9\u5e76\u975e\u4e00\u79cd\u53ef\u6301\u7eed\u7684\u65b9\u6cd5\u3002<\/p>\n<p>\u8981\u5728\u8fd9\u4e2a\u96c6\u7fa4\u4e0a\u5b9e\u73b0\u9ad8\u6027\u80fd\u8bad\u7ec3\uff0c\u5c31\u9700\u8981\u6bcf\u4e2a\u7ec4\u4ef6\uff08InfiniBand\u3001\u4ee5\u592a\u7f51\u3001GPU \u548c\u8282\u70b9\u672c\u8eab\uff09\u90fd\u8fd1\u4e4e\u5b8c\u7f8e\u5730\u5de5\u4f5c\u3002\u5982\u679c\u8fd9 12,000 \u591a\u4e2a\u8fde\u63a5\u4e2d\u6709\u4efb\u4f55\u4e00\u4e2a\u6709\u70b9\u4e0d\u7a33\u5b9a\uff0c\u5c31\u4f1a\u62d6\u6162\u6574\u4f53\u7684\u8bad\u7ec3\u8fd0\u884c\u901f\u5ea6\u3002<\/p>\n<p>\u672c\u6587\u63a5\u4e0b\u6765\u7684\u5185\u5bb9\u5c31\u662f\u4ecb\u7ecd\u5982\u4f55\u8ba9\u8fd9\u4e00\u5207\u90fd\u5b8c\u7f8e\u4e14\u7a33\u5b9a\u5730\u8fd0\u884c\u3002<\/p>\n<p><strong>\u8fc7\u7a0b\uff1a\u5982\u4f55\u5c06\u88f8\u673a\u53d8\u6210\u5b8c\u5168\u53ef\u8fd0\u884c\u7684\u96c6\u7fa4<\/strong><\/p>\n<p><strong>\u914d\u7f6e\u5404\u53f0\u673a\u5668<\/strong><\/p>\n<p>\u5728\u901a\u8fc7\u7ba1\u7406\u7f51\u7edc\u5efa\u7acb\u4e86\u4e0e\u96c6\u7fa4\u7684\u521d\u59cb\u4ee5\u592a\u7f51\u8fde\u63a5\u4e4b\u540e\uff0c\u5c31\u83b7\u5f97\u4e86\u57fa\u677f\u7ba1\u7406\u63a7\u5236\u5668\uff08BMC\/baseboard management controller\uff09\u7684\u8bbf\u95ee\u51ed\u8bc1\u3002BMC \u662f\u4e00\u79cd\u8fdc\u7a0b\u76d1\u63a7\u4e3b\u673a\u7cfb\u7edf\u7684\u4e13\u7528\u670d\u52a1\u5904\u7406\u5668\uff0c\u5e76\u4e14\u901a\u5e38\u8fde\u63a5\u5230\u4e00\u4e2a\u5206\u7acb\u7684\u7f51\u7edc\u3002\u5b83\u80fd\u8ba9\u6211\u4eec\u5c31\u50cf\u662f\u4eb2\u8eab\u5728\u73b0\u573a\u4e00\u6837\u64cd\u4f5c\u673a\u5668\uff0c\u5e76\u8fd8\u989d\u5916\u63d0\u4f9b\u4e86\u786c\u4ef6\u5065\u5eb7\u72b6\u51b5\u3001BIOS \u8bbe\u7f6e\u548c\u7535\u6e90\u7ba1\u7406\u7684 API\u3002<\/p>\n<p>\u914d\u5907\u597d\u8fd9\u4e9b\u7ec4\u4ef6\u540e\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u64b8\u8d77\u8896\u5b50\uff0c\u5f00\u59cb\u8bbe\u7f6e\u96c6\u7fa4\u4e86\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 0\uff1a\u5148\u914d\u7f6e\u597d\u4e00\u53f0\u673a\u5668<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u9996\u5148\u4f7f\u7528 iDRAC\uff08\u6234\u5c14\u7684\u57fa\u677f\u7ba1\u7406\u63a7\u5236\u5668\uff09\u5728\u4e00\u53f0\u670d\u52a1\u5668\u4e0a\u5b89\u88c5 Ubuntu 22.04\uff0c\u7136\u540e\u518d\u57fa\u4e8e\u8fd9\u4e2a<mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark>\u8bbe\u7f6e\u5176\u4ed6\u4e00\u5207\u3002iDRAC \u7684\u4e00\u9879\u80fd\u529b\u662f\u5141\u8bb8\u4ece\u672c\u5730\u8ba1\u7b97\u673a\u5b89\u88c5\u548c\u542f\u52a8 ISO \u955c\u50cf\uff0c\u5e76\u901a\u8fc7\u6d4f\u89c8\u5668\u63d0\u4f9b\u4e00\u4e2a\u865a\u62df\u63a7\u5236\u53f0\u3002\u7406\u60f3\u60c5\u51b5\u4e0b\uff0c\u8fd9\u662f\u8be5\u8fc7\u7a0b\u4e2d\u552f\u4e00\u7684\u624b\u52a8\u5b89\u88c5\u6b65\u9aa4\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 1\uff1a\u5728\u6bcf\u53f0\u673a\u5668\u4e0a\u5b89\u88c5<mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark><\/p>\n<\/li>\n<\/ul>\n<p>\u5728\u914d\u7f6e\u5b8c\u7b2c\u4e00\u53f0\u673a\u5668\u4e4b\u540e\uff0c\u7ee7\u7eed\u5b89\u88c5 Ubuntu \u7684 Metal-as-a-Service (MAAS) \u8f6f\u4ef6\u4ee5\u5e2e\u52a9\u914d\u7f6e\u5269\u4f59\u7684\u670d\u52a1\u5668\u3002\u4f7f\u7528\u9884\u542f\u52a8\u6267\u884c\u73af\u5883\u534f\u8bae\uff08PXE\uff09\u542f\u52a8\u548c\u81ea\u52a8\u5316 iDRAC \u5de5\u5177\uff0c\u53ef\u6307\u793a\u6bcf\u53f0\u673a\u5668\u4ece\u7f51\u7edc\u542f\u52a8\u5e76\u914d\u7f6e MAAS \u4ee5\u54cd\u5e94 PXE \u542f\u52a8\u8bf7\u6c42\u3002\u5728\u6267\u884c\u521d\u59cb\u7684\u7f51\u7edc\u542f\u52a8\u65f6\uff0c\u670d\u52a1\u5668\u4f1a\u901a\u8fc7\u52a8\u6001 IP \u5206\u914d\u534f\u8bae DHCP \u4ece MAAS \u83b7\u5f97\u4e00\u4e2a IP \u548c\u4e00\u4e2a\u521d\u59cb\u5185\u6838\uff0c\u800c\u65e0\u9700\u5728\u672c\u5730\u9a71\u52a8\u5668\u4e0a\u5b89\u88c5\u4efb\u4f55\u4e1c\u897f\u3002\u8fd9\u662f\u7528\u4e8e\u81ea\u52a8\u91cd\u590d\u6267\u884c<mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark>\u5b89\u88c5\u7684\u57fa\u672c\u73af\u5883\u3002\u4ece\u7406\u8bba\u4e0a\u8bb2\uff0c\u6211\u4eec\u53ea\u9700\u7b49\u5f85\u7b2c\u4e00\u6b21\u542f\u52a8\uff0c\u7136\u540e\u4e00\u5207\u90fd\u4f1a\u88ab\u5904\u7406\u597d\u3002\u4f46\u5b9e\u9645\u4e0a\uff0cMAAS \u4e0e BMC \u7684\u6574\u5408\u5e76\u4e0d\u53ef\u9760\uff0c\u56e0\u6b64\u6211\u4eec\u4f7f\u7528 iDRAC API \u6765\u4e8b\u5148\u6536\u96c6\u6bcf\u53f0\u673a\u5668\u7684 MAC \u5730\u5740\uff08\u4e00\u79cd\u552f\u4e00\u7684\u7269\u7406\u786c\u4ef6\u6807\u8bc6\u7b26\uff09\u3002<\/p>\n<p>\u5728\u8fd9\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cMAAS \u901a\u5e38\u662f\u690e\u6808\u4e2d\u6bd4\u8f83\u53ef\u9760\u7684\u7ec4\u4ef6\u3002\u4f46\u662f\uff0c\u6211\u4eec\u5728\u5f00\u59cb\u65f6\u9047\u5230\u4e86\u4e00\u4e9b\u6211\u4eec\u7684\u8bbe\u7f6e\u7279\u6709\u7684\u95ee\u9898\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5728\u914d\u7f6e\u524d\u51e0\u53f0\u673a\u5668\u65f6\uff0c\u7531\u4e8e\u65f6\u949f\u76f8\u5dee\u592a\u5927\uff0cHTTPS \u8bc1\u4e66\u9a8c\u8bc1\u95ee\u9898\u5bfc\u81f4\u65e0\u6cd5\u901a\u8fc7 apt \u5b89\u88c5\u4efb\u4f55\u4e1c\u897f\u3002\u4e0e\u6b64\u76f8\u5173\u7684\u662f\uff0c\u7531\u4e8e MAAS \u670d\u52a1\u5668\u5fc5\u987b\u8d1f\u8d23\u5f88\u591a\u4e8b\u60c5\uff08DHCP \u670d\u52a1\u5668\u3001\u7528\u4e8e\u5c06\u4e3b\u673a\u540d\u89e3\u6790\u6210 IP \u7684 DNS \u670d\u52a1\u5668\u3001\u4e3b\u673a\u548c\u5b98\u65b9 Ubuntu \u8f6f\u4ef6\u5305\u670d\u52a1\u5668\u4e4b\u95f4\u7684 HTTP \u4ee3\u7406\u3001NTP \u670d\u52a1\u5668\u3001cloud-init \u914d\u7f6e\u7ba1\u7406\u3001\u7528\u4e8e\u5c06 MAC \u5730\u5740\u8fde\u63a5\u5230 IP \u5230\u4e3b\u673a\u540d\u518d\u5230\u81ea\u5b9a\u4e49\u5143\u6570\u636e\u7684 ground truth <mark data-type=\"concepts\" data-id=\"700f9c0f-1e8b-4fde-8bae-6de39c13f022\">\u6570\u636e\u5e93<\/mark>\uff09\uff0c\u56e0\u6b64\u6211\u4eec\u5f88\u96be\u4ece\u6839\u6e90\u4e0a\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\u3002\u6b64\u5916\uff0c\u8fd8\u6709 MAAS \u914d\u7f6e\u751f\u547d\u5468\u671f\u7684<mark data-type=\"concepts\" data-id=\"356ac86d-6db9-47be-a5fb-b1c9bb405bcc\">\u5b66\u4e60\u66f2\u7ebf<\/mark>\u95ee\u9898\uff0c\u56e0\u4e3a\u662f\u8bbe\u8ba1\u76ee\u6807\u662f\u5904\u7406\u7ba1\u7406\u7eff\u5730\u90e8\u7f72\uff08greenfield deployment\uff09\u7684\u590d\u6742\u6027\u4ee5\u53ca\u8282\u70b9\u7684\u9010\u6b65\u8fc1\u79fb\u548c\u5404\u79cd\u8c03\u8bd5 \/ \u4e0d\u5065\u5eb7\u7684\u4e2d\u95f4\u72b6\u6001\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 2\uff1a\u8bca\u65ad\u635f\u574f\u7684\u673a\u5668<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u53d1\u73b0\u5927\u7ea6 10% \u7684\u673a\u5668\u90fd\u65e0\u6cd5\u542f\u52a8\uff0c\u4e3b\u8981\u539f\u56e0\u662f\u670d\u52a1\u5668\u7684\u7269\u7406\u95ee\u9898\u3002\u8fd9\u662f\u8bbe\u7f6e\u5927\u578b GPU \u96c6\u7fa4\u7684\u5e38\u89c1\u60c5\u51b5\u3002\u6211\u4eec\u9047\u5230\u7684\u60c5\u51b5\u5305\u62ec\uff1a\u6ca1\u63a5\u7f51\u7ebf\u6216\u63a5\u9519\u4e86\u7f51\u7ebf\u3001iDRAC \u4e2d\u7684\u786c\u4ef6\u95ee\u9898\u3001\u7535\u6e90\u5355\u5143\u635f\u574f\u3001NVME\uff08\u5feb\u901f\u975e\u6613\u5931\u6027\u5185\u5b58\uff09\u9a71\u52a8\u635f\u574f\u3001\u5185\u90e8\u7ebf\u8def\u7f3a\u5931\u3001\u7f51\u5361\u6216 GPU \u65e0\u6cd5\u663e\u793a\u3002\u6211\u4eec\u81ea\u52a8\u68c0\u67e5\u4e86\u8fd9\u4e9b\u95ee\u9898\uff0c\u5c06\u4e00\u4e9b\u673a\u5668\u9000\u56de\u7ed9\u4e86\u6234\u5c14\u4ee5\u91cd\u65b0\u6d4b\u8bd5\uff0c\u5e76\u4e3a\u6570\u636e\u4e2d\u5fc3\u5de5\u4f5c\u4eba\u5458\u63d0\u4ea4\u76f8\u5e94\u7684\u5de5\u5355\u3002\u6211\u4eec\u81ea\u5df1\u4e0a\u624b\u914d\u7f6e\u96c6\u7fa4\u7684\u4e00\u4e2a\u4f18\u52bf\u662f\uff1a\u5728\u7b49\u5f85\u7ef4\u62a4\u67d0\u4e9b\u673a\u5668\u7684\u540c\u65f6\u5c31\u80fd\u7acb\u5373\u4f7f\u7528\u5065\u5eb7\u7684\u673a\u5668\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 3\uff1a\u6700\u5c0f\u53ef\u884c\u53ef\u89c2\u5bdf\u673a\u5668<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u7ee7\u7eed\u5728\u6bcf\u53f0\u670d\u52a1\u5668\u4e0a\u8fdb\u884c\u5982\u4e0b\u8bbe\u7f6e\uff1a<\/p>\n<p>1.Docker\uff08\u4ee5\u4fbf\u66f4\u8f7b\u677e\u5730\u8fd0\u884c\u670d\u52a1\u548c\u8bad\u7ec3\u4f5c\u4e1a\uff09<\/p>\n<p>2. \u6570\u636e\u4e2d\u5fc3 GPU \u9a71\u52a8<\/p>\n<p>3.Prometheus \u8282\u70b9\u5bfc\u51fa\u5de5\u5177\uff08\u7528\u4e8e\u5bfc\u51fa\u7a33\u5b9a\u7684\u786c\u4ef6 \/ <mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark>\u6307\u6807\u6570\u636e\u6d41\uff09<\/p>\n<p>4.DCGM \u5bfc\u51fa\u5de5\u5177\uff08\u7528\u4e8e\u4ece\u82f1\u4f1f\u8fbe GPU \u5bfc\u51fa\u989d\u5916\u7684\u6307\u6807\u6570\u636e\uff0c\u5982 GPU \u72b6\u6001\u3001\u65f6\u949f\u3001\u5229\u7528\u7387\uff09<\/p>\n<p>5. \u6240\u6709\u975e<mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark>\u9a71\u52a8\u7684 RAIDZ ZFS \u6c60\uff0c\u8fd9\u8ba9\u673a\u5668\u5728\u67d0\u4e2a\u9a71\u52a8\u5931\u6548\u65f6\u4e5f\u80fd\u7ee7\u7eed\u5de5\u4f5c\uff0c\u540c\u65f6\u8fd8\u80fd\u514d\u8d39\u63d0\u4f9b\u900f\u660e\u7684\u538b\u7f29\uff08\u8fd9\u5bf9\u7eaf\u6587\u672c\u6570\u636e\u96c6\u548c\u91cd\u590d\u6027\u65e5\u5fd7\u5c24\u5176\u6709\u7528 \u2014\u2014 \u76f8\u6bd4\u4e8e\u4e0d\u4f7f\u7528\u8be5\u5de5\u5177\uff0c\u4f7f\u7528\u8be5\u5de5\u5177\u901a\u5e38\u80fd\u5c06\u53ef\u4f7f\u7528\u7684\u7a7a\u95f4\u589e\u5927 10 \u500d\uff09<\/p>\n<p>\u7136\u540e\u6211\u4eec\u8fd0\u884c\u57fa\u672c\u7684 GPU \u8bca\u65ad\u4ee5\u786e\u5b9a GPU \u662f\u5426\u5927\u4f53\u529f\u80fd\u6b63\u5e38 \u2014\u2014 \u4e0d\u6b63\u5e38\u7684\u901a\u5e38\u4f1a\u5728\u51e0\u4e2a\u5c0f\u65f6\u5185\u51fa\u73b0\u786c\u4ef6\u95ee\u9898\u3002<\/p>\n<p>\u5728\u6b64\u671f\u95f4\uff0c\u5f53\u6211\u4eec\u8bd5\u56fe\u540c\u65f6\u5728\u5168\u90e8 400 \u4e2a\u8282\u70b9\u4e0a\u5b89\u88c5\u8f6f\u4ef6\u5305\u65f6\uff0c\u6211\u4eec\u906d\u9047\u4e86\u5e26\u5bbd\u74f6\u9888\u3002\u8fd9\u662f\u6211\u4eec\u7b2c\u4e00\u6b21\u5728\u6570\u636e\u4e2d\u5fc3\u90e8\u7f72\u7684\u591a\u4e2a\u7ec4\u4ef6\u4e0a\u6536\u5230\u9ad8\u6e29\u8fc7\u70ed\u8b66\u62a5\u3002\u8fd9\u9996\u6279\u53d1\u70ed\u95ee\u9898\u57fa\u672c\u4e0a\u90fd\u901a\u8fc7\u56fa\u4ef6\u66f4\u65b0\u5f97\u5230\u4e86\u89e3\u51b3\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 4\uff1a\u5355\u8282\u70b9\u7684 GPU \u8bad\u7ec3<\/p>\n<\/li>\n<\/ul>\n<p>\u4e0b\u4e00\u6b65\u662f\u786e\u4fdd\u6bcf\u53f0\u673a\u5668\u90fd\u80fd\u591f\u5355\u72ec\u5904\u7406\u771f\u5b9e\u7684 GPU \u5de5\u4f5c\u8d1f\u8f7d\u3002\u5f88\u591a\u673a\u5668\u90fd\u65e0\u6cd5\u505a\u5230\u8fd9\u4e00\u70b9\uff0c\u95ee\u9898\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p>GPU \u76f8\u5173\u7684\u9519\u8bef\uff0c\u8fd9\u7c7b\u95ee\u9898\u57fa\u672c\u90fd\u53ef\u901a\u8fc7\u5c06 GPU \u5361\u91cd\u65b0\u63d2\u5165\u5361\u69fd\u6765\u89e3\u51b3\uff1a\u5c06 200 \u78c5\u91cd\u7684\u670d\u52a1\u5668\u4ece\u673a\u67b6\u4e0a\u6ed1\u51fa\u6765\uff0c\u79fb\u9664\u673a\u76d6\u548c GPU \u4e4b\u95f4\u7684\u6240\u6709\u7ebf\u7f06\uff0c\u7136\u540e\u53d6\u51fa GPU\uff0c\u518d\u91cd\u65b0\u88c5\u4e0a GPU\uff0c\u4e4b\u540e\u518d\u91cd\u65b0\u63a5\u4e0a\u7ebf\u7f06\u5e76\u628a\u670d\u52a1\u5668\u63a8\u56de\u673a\u67b6\u3002<\/p>\n<\/li>\n<li>\n<p>\u6839\u636e Ubuntu \u670d\u52a1\u5668\u65e5\u5fd7\uff0cGPU \u548c PCIe \u603b\u7ebf\u6216\u7f51\u5361\u4e4b\u95f4\u7684\u8bb8\u591a\u7ebf\u7f06\u90fd\u53d1\u51fa\u4e86\u8fd9\u6837\u7684\u62a5\u9519\uff1a\u300climited width: x4 &lt; x16\u300d\u3002\u5728\u66f4\u65b0 PCIe \u4ea4\u6362\u673a\u603b\u7ebf\u56fa\u4ef6\u540e\uff0c\u6211\u4eec\u53d1\u73b0\u5927\u7ea6\u56db\u5206\u4e4b\u4e00\u7684\u4e3b\u673a\u9700\u8981\u91cd\u65b0\u5b89\u88c5\u5185\u90e8 PCIe \u7ebf\u7f06 \u2014\u2014 \u5927\u6982\u662f\u56e0\u4e3a\u5916\u58f3\u548c GPU \u4e4b\u95f4\u7684\u7ebf\u7f06\u76f8\u5f53\u8106\u5f31\uff0c\u8fd9\u610f\u5473\u7740\u6bcf\u5f53\u5bf9 GPU \u8fdb\u884c\u7ef4\u62a4\u65f6\uff0c\u8fd9\u4e9b\u7ebf\u7f06\u90fd\u4f1a\u88ab\u63a8\u6324\u6216\u62d4\u6389\u3002<\/p>\n<\/li>\n<li>\n<p>\u8fd8\u6709\u4e00\u4e9b\u6742\u9879\u6545\u969c\u4e5f\u5f71\u54cd\u4e86\u51e0\u53f0\u4e3b\u673a\u3002\u6234\u5c14\u901a\u8fc7\u56fa\u4ef6\u5347\u7ea7\u5e2e\u52a9\u6211\u4eec\u89e3\u51b3\u4e86\u4e00\u4e9b\u95ee\u9898\uff1a<\/p>\n<\/li>\n<li>\n<p>NVMe \u9a71\u52a8\u5668\u6ca1\u6709\u663e\u793a\u6545\u969c\uff0c\u4f46\u89e6\u6478\u65f6\u4f1a\u9501\u5b9a\u6574\u53f0\u673a\u5668\u3002<\/p>\n<\/li>\n<li>\n<p>\u786c\u76d8\u9a71\u52a8\u5668\u5728 Linux \u4e0b\u4ee5\u968f\u673a\u987a\u5e8f\u663e\u793a\uff0c\u8fd9\u7ed9 MAAS \u9020\u6210\u4e86\u6df7\u4e71\uff0c\u5e76\u4f1a\u5bfc\u81f4<mark data-type=\"concepts\" data-id=\"9c0b81d8-32df-4e0c-a5f8-f02f69c55040\">\u64cd\u4f5c\u7cfb\u7edf<\/mark>\u88ab\u5b89\u88c5\u5728\u9519\u8bef\u7684\u9a71\u52a8\u5668\u4e0a\u3002<\/p>\n<\/li>\n<li>\n<p>\u6e29\u5ea6\u8bfb\u6570\u9519\u8bef\uff0c\u8fd9\u4f1a\u5bfc\u81f4\u98ce\u6247\u4e00\u76f4\u5168\u901f\u8fd0\u8f6c\u3002\u5176\u539f\u56e0\u53ef\u80fd\u662f\u82f1\u4f1f\u8fbe\u9a71\u52a8\u6709\u95ee\u9898\uff0c\u8fd9\u901a\u8fc7\u964d\u7ea7\u9a71\u52a8\u7248\u672c\u800c\u5f97\u5230\u4e86\u89e3\u51b3\u3002<\/p>\n<\/li>\n<li>\n<p>CPU \u7684\u52a8\u6001\u8c03\u9891\u5931\u63a7\uff0c\u5c06\u5de5\u4f5c\u5185\u6838\u9650\u5236\u4e3a 2 GHz\u3002<\/p>\n<\/li>\n<li>\n<p>\u76f4\u63a5\u7684 GPU-GPU \u901a\u4fe1\uff08GDR \u6216 GPUDirect RDMA Peer Memory Client\uff09\u65e0\u6cd5\u6210\u529f\u5e94\u7528\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u914d\u7f6e InfiniBand<\/strong><\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 0\uff1a\u5b89\u88c5 UFM<\/p>\n<\/li>\n<\/ul>\n<p>InfiniBand \u7684\u4e00\u4e2a\u4f18\u52bf\u662f\u5176\u4e2d\u5fc3\u5316\u7684\u8bbe\u8ba1\uff0c\u8fd9\u6837\u4e00\u6765\u6574\u4e2a\u7f51\u7edc\u5c31\u6709\u4e86\u4e00\u4e2a\u5927\u8111\u3002\u56e0\u6b64\uff0c\u5bf9\u4e8e\u6574\u4e2a\u7f51\u7edc\u7ed3\u6784\u4e2d\u7684 320 \u4e2a\u7f51\u7edc\u4ea4\u6362\u673a\uff0c\u6211\u4eec\u53ea\u9700\u5904\u7406\u5176\u4e2d\u4e00\u4e2a\u5b9e\u4f8b\u3002\u6211\u4eec\u7684\u9996\u4e2a\u4efb\u52a1\u662f\u641e\u6e05\u695a\u54ea\u53f0\u4ea4\u6362\u673a\u8fde\u63a5\u4e86\u54ea\u4e9b\u673a\u5668\uff0c\u7136\u540e\u5c06\u5176\u4e0e\u63a5\u7ebf\u56fe\u5173\u8054\u8d77\u6765\uff0c\u5e76\u6839\u636e\u4ea4\u6362\u673a\u7684\u7269\u7406\u4f4d\u7f6e\u91cd\u65b0\u547d\u540d\u5b83\u4eec\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 1\uff1a\u91cd\u65b0\u5e03\u7ebf<\/p>\n<\/li>\n<\/ul>\n<p>\u4e00\u5f00\u59cb\uff0cUFM \u65e0\u6cd5\u68c0\u6d4b\u5230\u90a3 320 \u53f0\u4ea4\u6362\u673a\uff0c\u66f4\u522b\u8bf4\u672c\u5e94\u51fa\u73b0\u5728\u7f51\u7edc\u7ed3\u6784\u4e2d\u7684\u4e3b\u673a\u4e86\u3002\u5728\u4e0e\u6211\u4eec\u7684\u6570\u636e\u4e2d\u5fc3\u5408\u4f5c\u4f19\u4f34\u5546\u8ba8\u4e4b\u540e\uff0c\u6211\u4eec\u786e\u8ba4\u8fd9\u4e9b\u4ea4\u6362\u673a\u5df2\u901a\u7535\u5e76\u63a5\u597d\u4e86\u7ebf\uff0c\u4f46\u4f9d\u7136\u65e0\u6cd5\u68c0\u6d4b\u5230\u3002\u5728\u68c0\u67e5\u7f51\u7edc\u5e03\u7ebf\u5217\u8868\u540e\uff0c\u6211\u4eec\u6ce8\u610f\u5230\u8be5\u7f51\u7edc\u7ed3\u6784\u7684\u9876\u5c42\u8bbe\u8ba1\u6709\u8bef\uff1a\u8fd9\u4e2a\u7ed3\u6784\u4e0d\u662f\u7edf\u4e00\u7684\uff0c\u800c\u662f\u5206\u6210\u4e86\u516b\u4e2a\u6ca1\u6709\u516c\u5171\u8def\u7531\u8def\u5f84\u7684\u4e92\u76f8\u8131\u79bb\u7684\u7f51\u7edc\u3002\u5728\u91cd\u65b0\u63a5\u7ebf\u4e4b\u540e\uff0c\u6211\u4eec\u6dfb\u52a0\u4e86\u68c0\u67e5\u6b65\u9aa4\uff0c\u4ee5\u9a8c\u8bc1\u6240\u6709\u7269\u7406\u8fde\u63a5\u662f\u5426\u4e0e\u65b0\u8bbe\u8ba1\u4e00\u81f4\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 2\uff1a\u4e00\u4e07\u6b21\u6e29\u5ea6\u544a\u8b66\uff08alert\uff09<\/p>\n<\/li>\n<\/ul>\n<p>\u5728\u89e3\u51b3\u4e86\u7269\u7406\u63a5\u7ebf\u95ee\u9898\u4e4b\u540e\uff0cInfiniBand \u6210\u529f\u5efa\u7acb\u4e86\u4e0e\u7f51\u7edc\u7ed3\u6784\u4e2d\u6240\u6709 InfiniBand \u4ea4\u6362\u673a\u7684\u8054\u7cfb\u3002\u4f46\u662f\uff0c\u51e0\u4e4e\u6bcf\u4e2a\u4ea4\u6362\u673a\u7aef\u53e3\u90fd\u5f00\u59cb\u62a5\u544a\u6e29\u5ea6\u8fc7\u9ad8\uff0c\u6709\u65f6\u5019\u8d85\u8fc7 70 \u2103\uff0c\u5373\u4fbf\u5b83\u4eec\u8fd8\u6ca1\u6709\u4f20\u8f93\u6570\u636e\u3002\u6211\u4eec\u53d1\u73b0\u8fd9\u4e2a\u95ee\u9898\u6e90\u81ea\u540c\u4e00\u673a\u67b6\u4e2d\u4ea4\u6362\u673a\u4e4b\u95f4\u7684\u5f00\u653e\u7a7a\u95f4\uff0c\u8fd9\u5bfc\u81f4\u70ed\u7a7a\u6c14\u56de\u6d41\u5230\u4e86\u524d\u9762\u3002\u6211\u4eec\u7684\u6570\u636e\u4e2d\u5fc3\u5408\u4f5c\u4f19\u4f34\u5e2e\u52a9\u6211\u4eec\u5feb\u901f\u8bca\u65ad\u51fa\u4e86\u8be5\u95ee\u9898\u5e76\u5236\u5b9a\u4e86\u5408\u9002\u7684\u89e3\u51b3\u65b9\u6cd5\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 3\uff1a1800 \u6b21\u544a\u8b66<\/p>\n<\/li>\n<\/ul>\n<p>\u8bb8\u591a\u7aef\u53e3\u8fd8\u6709\u5f88\u9ad8\u7684\u9519\u8bef\u7387\uff0c\u6216\u5728\u6b63\u5e38\u548c\u635f\u574f\u72b6\u6001\u4e4b\u95f4\u6765\u56de\u53d8\u52a8\uff0c\u8fd9\u88ab\u79f0\u4e3a\u300c\u6296\u52a8\uff08flapping\uff09\u300d\u3002\u8fd9\u4e9b\u95ee\u9898\u53ea\u6709\u5728\u5b9e\u9645\u4f7f\u7528\u8fd9\u4e9b\u7aef\u53e3\u65f6\u624d\u4f1a\u51fa\u73b0\uff0c\u6240\u4ee5\u5f88\u96be\u9884\u5148\u68c0\u6d4b\uff0c\u56e0\u4e3a\u6211\u4eec\u7684\u6574\u4e2a\u7ed3\u6784\u7531 10,000 \u6761\u9ad8\u5ea6\u5197\u4f59\u7684\u94fe\u8def\u7ec4\u6210\u3002\u6211\u4eec\u7684\u6570\u636e\u4e2d\u5fc3\u5408\u4f5c\u4f19\u4f34\u5e2e\u52a9\u6e05\u6d01\u548c\u91cd\u65b0\u5b89\u88c5\u544a\u8b66\u7684\u7aef\u53e3\uff0c\u6211\u4eec\u5728\u7b49\u5f85\u66ff\u6362\u65f6\u7981\u7528\u4e86\u5269\u4f59\u7684\u8b66\u62a5\u6536\u53d1\u5668\u3002<\/p>\n<p>\u5c3d\u7ba1 InfiniBand \u80fd\u5f39\u6027\u5730\u5e94\u5bf9\u786c\u4ef6\u6545\u969c\uff0c\u4f46\u4e00\u65e6\u5927\u7ea6 10% \u7684\u7ed3\u6784\u5f00\u59cb\u51fa\u73b0\u95ee\u9898\uff0c\u81ea\u9002\u5e94\u8def\u7531\u7b49\u529f\u80fd\u5c31\u65e0\u6cd5\u53ef\u9760\u5730\u8fd0\u884c\uff0c\u65e0\u6cd5\u89e3\u51b3\u5076\u5c14\u4e22\u5931\u94fe\u8def\u7684\u95ee\u9898\u3002<\/p>\n<p>\u5728\u6b64\u671f\u95f4\uff0c\u6211\u4eec\u6210\u529f\u4f7f\u7528 100 \u5230 200 \u53f0\u673a\u5668\u8fd0\u884c\u4e86\u591a\u8282\u70b9\u8bad\u7ec3\u3002\u6211\u4eec\u7684\u6d41\u7a0b\u6bd4\u8f83\u5373\u5174\uff1a\u6211\u4eec\u6709\u65f6\u4f1a\u968f\u673a\u542f\u52a8\u4e00\u7ec4\u8282\u70b9\uff0c\u89c2\u5bdf\u5b83\u4eec\u7684\u6027\u80fd\uff0c\u7136\u540e\u5c3d\u529b\u8ba9\u5176\u4e2d\u5c3d\u53ef\u80fd\u591a\u7684\u8282\u70b9\u4fdd\u6301\u8fd0\u884c\u3002\u8be5\u65b9\u6cd5\u53ef\u8ba9\u6211\u4eec\u627e\u5230\u8be5 InfiniBand \u7f51\u7edc\u7ed3\u6784\u4e2d\u4e00\u7ec4\u53ef\u9760\u7684\u5b50\u96c6\uff0c\u4f46\u96be\u5ea6\u5374\u5f88\u5927\uff0c\u56e0\u4e3a\u6bcf\u6b21\u90fd\u9700\u8981\u6539\u53d8\u7528\u4e8e\u8bad\u7ec3\u7684\u8282\u70b9\u96c6\u5408\uff0c\u7531\u6b64\u6539\u53d8\u9ed8\u8ba4\u7684 InfiniBand \u94fe\u8def\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 4\uff1aInfiniBand \u75af\u72c2\u71c3\u70e7<\/p>\n<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u66f4\u9ad8\u6548\u5730\u8bca\u65ad InfiniBand \u95ee\u9898\uff0c\u6211\u4eec\u4e13\u95e8\u4e3a\u6574\u4e2a\u96c6\u7fa4\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u5176\u4f5c\u7528\u662f\u540c\u65f6\u5c06\u5c3d\u53ef\u80fd\u591a\u7684\u6570\u636e\u63a8\u9001\u7ecf\u8fc7\u6574\u4e2a\u7ed3\u6784\u4e2d\u7684\u6bcf\u4e2a\u7aef\u53e3\u3002\u8fd9\u4e0d\u540c\u4e8e\u5728\u6574\u4e2a\u96c6\u7fa4\u4e0a\u8fd0\u884c\u4e00\u4e2a\u5927\u578b\u7684 all-reduce \u5de5\u4f5c\u8d1f\u8f7d\uff0c\u8fd9\u9700\u8981\u4f7f\u7528 NCCL \u6765\u4f18\u5316\u5404\u4e2a\u8282\u70b9\u4e4b\u4e2d\u7684\u901a\u4fe1\uff0c\u65b9\u6cd5\u662f\u4f7f\u7528 NVLink \u7ecf\u7531 Server PCIe Module (SXM) \u63d2\u69fd\u6765\u5b9e\u73b0 GPU \u901a\u4fe1\u3002<\/p>\n<p>\u76f8\u53cd\uff0c\u6211\u4eec\u9009\u62e9\u4e86\u4e00\u79cd\u86ee\u529b\u65b9\u6cd5\uff0c\u5e76\u8f7b\u677e\u53d6\u5f97\u4e86\u6210\u529f\u3002UFM \u4f1a\u5728\u5927\u591a\u6570\u7aef\u53e3\u7684\u6570\u636e\u4f20\u8f93\u91cf\u8d85\u8fc7\u7406\u8bba\u5bb9\u91cf\u7684 97% \u65f6\u5f00\u59cb\u53d1\u51fa\u8b66\u62a5\uff0c\u540c\u65f6\u4e00\u4e9b\u4ea4\u6362\u673a\u4f1a\u6682\u65f6\u5b95\u673a\u3002\u6211\u4eec\u8ba4\u4e3a\u80fd\u575a\u6301\u5230\u5f53\u5929\u7ed3\u675f\u65f6\u7684\u6bcf\u4e2a\u7aef\u53e3\u90fd\u662f\u8db3\u591f\u7a33\u5065\u7684\uff0c\u5176\u4f59\u7684\u90fd\u88ab\u7981\u7528\u6216\u79fb\u9664\u4ee5\u5f85\u7ef4\u4fee\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 5\uff1aGPUDirect RDMA<\/p>\n<\/li>\n<\/ul>\n<p>\u8981\u8ba9 GPU \u901a\u4fe1\u65f6\u4e0d\u4ea7\u751f CPU \u8ba1\u7b97\u5f00\u9500\uff0c\u6211\u4eec\u542f\u7528\u4e86\u4e00\u4e2a\u540d\u4e3a GPUDirect RDMA \u7684\u529f\u80fd\uff0c\u5176\u5141\u8bb8 InfiniBand \u7f51\u5361\u4e4b\u95f4\u76f4\u63a5\u901a\u4fe1\u3002\u8fd9\u6d89\u53ca\u4e24\u4e2a\u5173\u952e\u6b65\u9aa4\uff1a<\/p>\n<p>1. \u542f\u52a8\u4e00\u4e2a\u989d\u5916\u7684\u6838\u6a21\u5757<\/p>\n<p>2. \u786e\u4fdd PCIe Access Control Service (ACS) \u88ab\u7981\u7528\uff0c\u4ee5\u9632\u6b62 immediate hangs\uff08\u7acb\u5373\u6302\u8d77\uff09<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 6\uff1a\u6269\u589e\u300c\u9ec4\u91d1\u300d\u670d\u52a1\u5668<\/p>\n<\/li>\n<\/ul>\n<p>\u8981\u4f7f\u7528\u6700\u65b0\u786c\u4ef6\u6784\u5efa GPU \u96c6\u7fa4\uff0c\u4e00\u4e2a\u7ecf\u9a8c\u6cd5\u5219\u662f\uff1a\u6bcf\u5468\u90fd\u6709\u5927\u7ea6 3% \u7684\u673a\u5668\u51fa\u95ee\u9898\uff0c\u8981\u505a\u597d\u51c6\u5907\u3002<\/p>\n<p>\u4f46\u662f\uff0c\u9700\u8981\u8bf4\u660e\u4e00\u70b9\uff1a\u5e76\u4e0d\u662f\u6bcf\u53f0\u673a\u5668\u90fd\u7edf\u4e00\u6709 3% \u7684\u51e0\u7387\u53d1\u751f\u6545\u969c\uff0c\u800c\u662f\u5c11\u91cf\u4e0d\u5bf9\u4ed8\u7684\u673a\u5668\u53cd\u590d\u51fa\u73b0\u5404\u79cd\u4e0d\u540c\u95ee\u9898\uff0c\u76f4\u5230\u5c06\u5b83\u4eec\u59a5\u5584\u4fee\u590d\u3002\u8fd9\u5c31\u51f8\u663e\u4e86\u5728\u540c\u4e00\u7f51\u7edc\u7ed3\u6784\u4e2d\u914d\u5907\u5927\u91cf\u673a\u5668\u7684\u4f18\u52bf\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u7684\u505a\u6cd5\u4e0d\u662f\u968f\u4fbf\u627e\u4e9b\u673a\u5668\u6765\u8fd0\u884c\u5927\u89c4\u6a21\u8bad\u7ec3\uff0c\u5c31\u50cf\u6253\u5730\u9f20\u4e00\u6837\u770b\u54ea\u4e9b\u51fa\u95ee\u9898\uff0c\u800c\u662f\u4e13\u6ce8\u4e8e\u6269\u589e\u5df2\u77e5\u53ef\u9760\u7684\u670d\u52a1\u5668\uff0c\u4e5f\u5c31\u662f\u300c\u9ec4\u91d1\u300d\u670d\u52a1\u5668\u3002<\/p>\n<ul>\n<li>\n<p>\u6b65\u9aa4 7\uff1a\u7ef4\u62a4<\/p>\n<\/li>\n<\/ul>\n<p>InfiniBand \u7684\u7ef4\u62a4\u4e3b\u8981\u6d89\u53ca\u5230\u54cd\u5e94 UFM \u8b66\u62a5\u3001\u66f4\u6362\u6545\u969c\u7ebf\u7f06\u548c\u6536\u53d1\u5668\uff0c\u4ee5\u53ca\u5076\u5c14\u8bca\u65ad\u66f4\u56f0\u96be\u7684\u9519\u8bef\uff08\u6bd4\u5982\u4ea4\u6362\u673a\u6545\u969c\uff09\u3002\u5bfc\u81f4\u5927\u89c4\u6a21\u7ef4\u62a4\u7684\u56e0\u7d20\u901a\u5e38\u6709\u4e24\u4e2a\uff1a<\/p>\n<p>1. \u56fa\u4ef6\u66f4\u65b0\uff0c\u5c24\u5176\u662f\u96c6\u7fa4\u4e2d\u4ec5\u6709\u4e00\u534a\u5b8c\u6210\u66f4\u65b0\u65f6\uff0c\u8fd9\u53ef\u80fd\u5bfc\u81f4 UFM \u72b6\u6001\u635f\u574f\u5e76\u5fc5\u9700\u91cd\u542f\u6240\u6709 InfiniBand \u4ea4\u6362\u673a\u4e0a\u7684 UFM\u3002<\/p>\n<p>2.GPU \u76d2\u540c\u65f6\u5927\u89c4\u6a21\u91cd\u542f\uff0c\u8fd9\u53ef\u80fd\u4f1a\u5411 UFM \u72b6\u6001\u704c\u5165\u5927\u91cf\u66f4\u65b0\uff0c\u5e76\u540c\u6837\u9700\u8981\u91cd\u542f UFM \u670d\u52a1\u3002<\/p>\n<p><strong>\u786e\u4fdd\u673a\u5668\u5b8c\u5168\u5065\u5eb7<\/strong><\/p>\n<p>\u5728\u6b64\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u53d1\u73b0\u4e86\u5355\u53f0\u673a\u5668\u7684\u591a\u79cd\u6545\u969c\u6216\u51cf\u6162\u8bad\u7ec3\u901f\u5ea6\u7684\u65b9\u5f0f\u3002\u5176\u4e2d\u8bb8\u591a\u6545\u969c\u6a21\u5f0f\u5e76\u4e0d\u4f1a\u7acb\u5373\u663e\u73b0\uff0c\u56e0\u6b64\u6211\u4eec\u7f16\u5199\u4e86\u8bb8\u591a\u5065\u5eb7\u68c0\u67e5\u811a\u672c\uff0c\u4ee5\u68c0\u67e5\u4e3b\u673a\u662f\u5426\u8db3\u591f\u5065\u5eb7\u3002\u6211\u4eec\u5728\u8fd9\u91cc\u53d1\u5e03\u4e86\u8fd9\u4e9b\u4ee3\u7801\uff1ahttps:\/\/github.com\/imbue-ai\/cluster-health<\/p>\n<p>\u8bf7\u6ce8\u610f\uff0c\u8fd9\u4e9b\u5065\u5eb7\u68c0\u67e5\u4e2d\u7684\u5f88\u591a\u90fd\u7279\u5b9a\u4e8e\u6211\u4eec\u7684\u8fd0\u884c\u65f6\u73af\u5883\uff0c\u5e76\u4e0d\u4e00\u5b9a\u4e0e\u57fa\u7840\u786c\u4ef6\u76f8\u5173\uff0c\u4e5f\u4e0d\u4e00\u5b9a\u5bb9\u6613\u4fee\u590d\u6216\u81ea\u52a8\u5316\u3002\u8fd9\u662f\u8bbe\u8ba1\u51b3\u5b9a\u7684\uff1a\u4e3a\u4e86\u5b9e\u73b0\u8ba9\u6211\u4eec\u7684\u673a\u5668\u51c6\u5907\u597d\u8bad\u7ec3\u7684\u603b\u4f53\u76ee\u6807\uff0c\u6211\u4eec\u60f3\u8981\u4e00\u4e2a\u53ef\u4ee5\u76f4\u63a5\u4e86\u5f53\u5730\u56de\u7b54\u300c\u662f\u300d\u6216\u300c\u5426\u300d\u7684\u5355\u4e00\u5165\u53e3\u70b9\uff0c\u5e76\u4e14\u53ef\u4ee5\u6982\u62ec\u603b\u7ed3\u4efb\u610f\u6570\u91cf\u7684\u7ec6\u5fae\u7ec6\u8282\u3002<\/p>\n<ul>\n<li>\n<p>GPU \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86 GPU \u6570\u91cf\u662f\u5426\u6b63\u786e\u3001ECC\uff08\u9519\u8bef\u66f4\u6b63\u4ee3\u7801\uff09\u68c0\u67e5\u662f\u5426\u5df2\u542f\u7528\u4ee5\u53ca\u662f\u5426\u5b58\u5728 ECC \u9519\u8bef\u3002\u6211\u4eec\u8fd8\u68c0\u67e5\u4e86 NVLink \u62d3\u6251\uff08\u5c06 GPU \u5f7c\u6b64\u8fde\u63a5\u8d77\u6765\uff09\u662f\u5426\u5df2\u542f\u52a8\u4e14\u65e0\u9519\u8bef\u3002<\/p>\n<ul>\n<li>\n<p>\u78c1\u76d8\u7a7a\u95f4\u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86\u4e3b\u673a\u7684\u78c1\u76d8\u7a7a\u95f4\u5229\u7528\u7387\u662f\u5426\u8d85\u8fc7 95%\u3002<\/p>\n<ul>\n<li>\n<p>Docker \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86 Docker \u80fd\u5426\u5728\u8fde\u63a5\u4e86 GPU \u7684\u60c5\u51b5\u4e0b\u8fd0\u884c\u5bb9\u5668\uff08\u5373 NVIDIA Container Runtime \u662f\u5426\u6b63\u5e38\u5de5\u4f5c\uff09\uff0c\u8fd8\u68c0\u67e5\u4e86\u4e0e\u76d1\u63a7 \/ \u5206\u6790\u76f8\u5173\u7684 Docker \u5bb9\u5668\u662f\u5426\u5df2\u6fc0\u6d3b\u5e76\u83b7\u5f97\u4e86\u6b63\u786e\u7684\u4e3b\u673a\u6743\u9650\u3002<\/p>\n<ul>\n<li>\n<p>Dmesg \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86 dmesg \u4e2d\u662f\u5426\u5b58\u5728\u786c\u4ef6 Xids \u6216 SXid \u9519\u8bef\uff08\u7531 NVIDIA GPU \u6216 GPU \u95f4 NVIDIA \u4ea4\u6362\u673a\u5f15\u53d1\u7684\u6545\u969c\uff09\u3002\u6211\u4eec\u8fd8\u8bfb\u53d6\u4e86\u6240\u6709 dmesg \u65e5\u5fd7\u884c\uff0c\u4ee5\u9a8c\u8bc1\u5b83\u4eec\u662f\u5426\u90fd\u53ef\u4ee5\u5f52\u7c7b\u5230\u300c\u5e38\u89c1 \/ \u9884\u671f\u65e5\u5fd7\u884c\u300d\u5217\u8868\u4e2d\u3002<\/p>\n<ul>\n<li>\n<p>iDRAC \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86\u673a\u5668\u4e0a\u7684 iDRAC \u9519\u8bef\uff0c\u5176\u4e2d\u5ffd\u7565\u4e86\u975e\u81f4\u547d\u9519\u8bef\u6d88\u606f\u3002\u8fd9\u662f\u6234\u5c14\u8ba1\u7b97\u673a\u7279\u6709\u7684\u68c0\u67e5\uff0c\u6240\u4ee5\u6ca1\u6709\u88ab\u5305\u62ec\u5728\u6211\u4eec\u5f00\u6e90\u7684\u4ee3\u7801\u4e2d\u3002<\/p>\n<ul>\n<li>\n<p>\u78c1\u76d8\u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86 zpool \u662f\u5426\u5df2\u5b89\u88c5\uff0cDocker \u662f\u5426\u5df2\u6b63\u786e\u8fde\u63a5\u5230\u5b83\uff0c\u4ee5\u53ca\u5b83\u662f\u5426\u771f\u7684\u53ef\u4ee5\u5728\u4e0d\u9501\u5b9a CPU \u7684\u60c5\u51b5\u4e0b\u89e6\u53ca\u5b83\u3002<\/p>\n<ul>\n<li>\n<p>InfiniBand \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86 InfiniBand \u7684\u9519\u8bef\u7387\u662f\u5426\u4f1a\u589e\u52a0\u548c \/ \u6216\u9a71\u52a8\u56fa\u4ef6\u662f\u5426\u8fc7\u65f6\u3002<\/p>\n<ul>\n<li>\n<p>Nvlink \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86\u673a\u5668\u4e0a\u7684 NVLink \u9519\u8bef\u3002\u5b9e\u8df5\u4e2d\u770b\uff0c\u8fd9\u4f3c\u4e4e\u4e0d\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u5931\u8d25\uff0c\u4f46\u53ef\u80fd\u4f1a\u964d\u4f4e\u8bad\u7ec3\u901f\u5ea6\u3002<\/p>\n<ul>\n<li>\n<p>GDR \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86\u673a\u5668\u4e0a\u7684 GDR \u662f\u5426\u5df2\u542f\u7528\u3002<\/p>\n<ul>\n<li>\n<p>VBIOS \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u68c0\u67e5\u4e86 GPU \u7684 VBIOS \u7248\u672c\u4ee5\u53ca H100 \u57fa\u677f\u56fa\u4ef6\u662f\u5426\u662f\u6700\u65b0\u7684\u3002<\/p>\n<ul>\n<li>\n<p>Flint \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u4f7f\u7528 flint \u548c hca_self_test \u68c0\u67e5\u4e86 Mellanox OFED \u9a71\u52a8\u3001\u7f51\u5361\u56fa\u4ef6\u548c\u6536\u53d1\u5668\u56fa\u4ef6\u7684\u7248\u672c\u662f\u5426\u6b63\u786e\uff0c\u4ee5\u53ca\u5b83\u4eec\u662f\u5426\u9488\u5bf9\u82f1\u4f1f\u8fbe\u9a71\u52a8\u8fdb\u884c\u4e86\u6b63\u786e\u7f16\u8bd1\u3002<\/p>\n<ul>\n<li>\n<p>PSB \u5065\u5eb7\u68c0\u67e5<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark>\u4e86 PCIe \u8bbe\u5907\uff0c\u4ee5\u68c0\u67e5 GPU\u3001PSB\uff08PCIe \u4ea4\u6362\u673a\u603b\u7ebf\uff09\u548c\u7f51\u5361\u4e4b\u95f4\u7684\u8fde\u63a5\u901f\u5ea6\u548c\u5bbd\u5ea6\u662f\u5426\u7b26\u5408\u6211\u4eec\u7684\u9884\u671f\u3002\u6211\u4eec\u8fd8\u68c0\u67e5\u4e86\u4ea4\u6362\u673a\u56fa\u4ef6\u662f\u5426\u4e3a\u6700\u65b0\u7248\u672c\u3002\u8be5\u811a\u672c\u7531\u6234\u5c14\u800c\u975e Imbue \u5f00\u53d1\uff0c\u6240\u4ee5\u6211\u4eec\u76ee\u524d\u65e0\u6cd5\u5171\u4eab\u5b83\u3002<\/p>\n<p>\u9664\u4e86\u8fd9\u4e9b\u5feb\u901f\u5065\u5eb7\u68c0\u67e5\uff0c\u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u4e00\u4e9b\u66f4\u590d\u6742\u7684\u5065\u5eb7\u68c0\u67e5\uff0c\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p>\u901a\u8fc7 PyTorch \u521d\u59cb\u5316\u77e9\u9635\u8ba1\u7b97\uff0c\u4ee5\u53ca\u6d4b\u91cf NVLink \u5e26\u5bbd\u548c GPU \u8ba1\u7b97\u901f\u5ea6\u548c\u5185\u5b58\u3002\u6211\u4eec\u8bbe\u7f6e\u4e86\u9002\u5f53\u7684 GDR \u6807\u5fd7\u6765\u6d4b\u8bd5 InfiniBand \u548c NVLink\u3002<\/p>\n<\/li>\n<li>\n<p>\u4f7f\u7528 ib_write_bw \u548c \u2013use_cuda \u901a\u8fc7 IB \u5361\u53d1\u9001\u6570\u636e\u5e76\u6d4b\u91cf PCIe \u548c InfiniBand \u5361\u5e26\u5bbd\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u6301\u7eed\u4e86\u8f83\u957f\u65f6\u95f4\uff08\u7ea6 15 \u5206\u949f\uff09\uff0c\u4ee5\u786e\u4fdd\u80fd\u627e\u51fa\u6296\u52a8\u7684 InfiniBand \u94fe\u8def\u3002<\/p>\n<\/li>\n<li>\n<p>\u8fd0\u884c\u591a\u8282\u70b9\u8bca\u65ad\u8fd0\u884c\u4ee5\u68c0\u67e5 NCCL \u521d\u59cb\u5316\u80fd\u529b\u4ee5\u53ca\u5b83\u662f\u5426\u4f1a\u968f\u673a\u505c\u987f\u3002\u5982\u6709\u505c\u987f\uff0c\u5219\u6211\u4eec\u7684\u5206\u53c9\u7248 NCCL \u4ee3\u7801\u4f1a\u6dfb\u52a0\u989d\u5916\u7684\u65e5\u5fd7\u8bb0\u5f55\u3002\u8fd9\u9700\u8981 12 \u5230 24 \u5c0f\u65f6\u624d\u80fd\u68c0\u6d4b\u5230\u95ee\u9898\uff0c\u56e0\u6b64\u6211\u4eec\u901a\u5e38\u53ea\u5bf9\u65b0\u8282\u70b9\u6216\u6211\u4eec\u6000\u7591\u5b58\u5728\u95ee\u9898\u65f6\u8fd0\u884c\u6b64\u64cd\u4f5c\u3002<\/p>\n<\/li>\n<li>\n<p>\u68c0\u67e5 DCGM \u5bfc\u51fa\u662f\u5426\u5b58\u5728\u4efb\u4f55 GPU \u65f6\u949f\u8282\u6d41\u4e8b\u4ef6\uff08\u4e0d\u5305\u62ec\u9884\u671f\u7684 gpu_idle \u548c power_cap\uff09\u3002\u4e3a\u4e86\u68c0\u67e5\u8fd9\u4e9b\u7535\u6e90\u4e8b\u4ef6\uff0c\u6700\u597d\u7684\u65b9\u6cd5\u662f\u8fd0\u884c\u591a\u8282\u70b9\u8bad\u7ec3\uff0c\u4ee5\u540c\u65f6\u68c0\u67e5\u6240\u6709 GPU\u3001InfiniBand \u5361\u4ee5\u53ca CPU \u548c\u78c1\u76d8\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u8bca\u65ad\u5e38\u89c1\u7684\u8bad\u7ec3\u95ee\u9898<\/strong><\/p>\n<p>\u4e00\u65e6\u786c\u4ef6\u80fd\u6b63\u5e38\u5de5\u4f5c\uff0c\u5c31\u53ef\u4ee5\u5f00\u59cb\u8bad\u7ec3\u4e86\u3002<\/p>\n<p>\u8fd9\u4e00\u8282\u5c06\u57fa\u4e8e\u6211\u4eec\u5728\u6211\u4eec\u7684\u96c6\u7fa4\u4e0a\u8fd0\u884c\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u8bad\u7ec3\u7684\u7ecf\u9a8c\uff0c\u5206\u4eab\u4e00\u4e9b\u5177\u4f53\u7684\u8c03\u8bd5\u6b65\u9aa4\u548c\u89c1\u89e3\u3002<\/p>\n<ul>\n<li>\n<p>\u542f\u52a8\u65f6\u5d29\u6e83<\/p>\n<\/li>\n<\/ul>\n<p>\u4ece\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u8bb2\uff0c\u8fd9\u662f\u6240\u80fd\u9047\u5230\u7684\u6700\u597d\u7684\u9519\u8bef\uff0c\u56e0\u4e3a\u5176\uff08\u7406\u8bba\u4e0a\uff09\u5f88\u5bb9\u6613\u91cd\u73b0\u548c\u8fed\u4ee3\u3002<\/p>\n<p>\u6211\u4eec\u9996\u5148\u68c0\u67e5\u4e86\u6211\u4eec\u7684\u4ee3\u7801\u662f\u5426\u5728\u6b63\u786e\u7684\u7248\u672c\u3001\u914d\u7f6e\u548c\u73af\u5883\u53d8\u91cf\u4e0a\u8fd0\u884c\u3002\u867d\u7136\u5f88\u57fa\u7840\uff0c\u4f46\u6211\u4eec\u53d1\u73b0\u8fd9\u4e00\u70b9\u81f3\u5173\u91cd\u8981\uff1a\u786e\u4fdd\u542f\u52a8\u8bad\u7ec3\u8fc7\u7a0b\u662f\u53ef\u590d\u73b0\u4e14\u5bb9\u6613\u68c0\u67e5\u7684\u3002\u4e00\u5927\u539f\u56e0\u662f Docker \u955c\u50cf\u7f13\u5b58\u6216\u4e0d\u900f\u660e\u79d8\u5bc6\u914d\u7f6e\u7b49\u4e2d\u95f4\u62bd\u8c61\u53ef\u80fd\u4f1a\u9020\u6210\u6df7\u6dc6\u3002<\/p>\n<p>\u6211\u4eec\u6267\u884c\u7684\u53e6\u4e00\u4e2a\u57fa\u7840\u68c0\u67e5\u662f\u786e\u4fdd\u6240\u6709\u673a\u5668\u90fd\u5728\u7ebf\uff0c\u5e76\u4e14\u53ef\u4ee5\u8f7b\u677e\u5730\u805a\u5408\u548c\u68c0\u67e5\u6240\u53d1\u51fa\u7684\u6808\u8ddf\u8e2a\u8bb0\u5f55\u6216\u65e5\u5fd7\u3002\u6211\u4eec\u4f7f\u7528\u4e86 Loki\u3001Prometheus \u548c Grafana \u8f6f\u4ef6\u6808\uff0c\u4f46\u4efb\u4f55\u5408\u9002\u7684\u65e5\u5fd7\u805a\u5408\u6216\u8ddf\u8e2a SaaS \u7684\u5de5\u5177\u90fd\u53ef\u4ee5\u3002\u7531\u4e8e\u8fd9\u4e9b\u8bad\u7ec3\u8fd0\u884c\u8fc7\u7a0b\u672c\u8d28\u4e0a\u662f\u540c\u6b65\u7684\u548c\u5206\u5e03\u5f0f\u7684\uff0c\u56e0\u6b64\u7b2c\u4e00\u4e2a\u9519\u8bef\u5f80\u5f80\u5c31\u4f1a\u5bfc\u81f4\u4e00\u8fde\u4e32\u7684\u4e0d\u76f8\u5173\u9519\u8bef\u3002\u5728\u8fd9\u91cc\uff0c\u5065\u5eb7\u68c0\u67e5\u8fd8\u53ef\u4ee5\u5e2e\u52a9\u7acb\u9a6c\u68c0\u6d4b\u51fa\u786c\u76d8\u9a71\u52a8\u5668\u635f\u574f\u6216 GPU \u7f3a\u5931\u6216\u65e0\u6548\u7b49\u9519\u8bef\u3002<\/p>\n<p>\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u5728\u53d1\u751f\u6545\u969c\u65f6\u81ea\u52a8\u91cd\u542f\u7684\u7cfb\u7edf\uff0c\u8fd9\u4f7f\u5f97\u65e5\u5fd7\u548c\u9519\u8bef\u805a\u5408\u53d8\u5f97\u66f4\u52a0\u91cd\u8981\uff0c\u4ee5\u907f\u514d\u6df7\u6dc6\u6765\u81ea\u4e0d\u540c\u91cd\u542f\u7684\u9519\u8bef\u3002\u6211\u4eec\u9047\u5230\u7684\u4e00\u4e9b\u5e38\u89c1\u9519\u8bef\u5305\u62ec\uff1a<\/p>\n<p>1.\u300cForward order differs across ranks: rank 0 is all-gathering 43 parameters while rank 1228 is all-gathering 1 parameters\u300d\u8fd9\u6837\u7684\u9519\u8bef\u3002\u6211\u4eec\u53d1\u73b0\u8fd9\u662f PyTorch \u5b8c\u5168\u5206\u7247\u6570\u636e\u5e76\u884c\uff08FSDP\uff09\u5b9e\u73b0\u7684\u4e00\u4e2a\u5947\u602a\u7279\u6027\uff0c\u53ef\u901a\u8fc7\u91cd\u542f\u89e3\u51b3\u3002<\/p>\n<p>2.GPU \u5185\u5b58\u4e0d\u8db3\uff08OOM\uff09\u9519\u8bef\uff0c\u770b\u8d77\u6765\u50cf\u8fd9\u6837\uff1a\u300cCUDA out of memory. Tried to allocate \u2026\u300d\u901a\u8fc7\u591a\u6b21\u68c0\u67e5\u6211\u4eec\u7684\u914d\u7f6e\u548c\u4ee3\u7801\u5e76\u64a4\u9500\u8fd1\u671f\u7684\u4ee3\u7801\u4fee\u6539\uff08\u7531\u4e8e\u542f\u52a8\u671f\u95f4 PyTorch \u8bbe\u5907\u89c4\u683c\u4e0d\u6b63\u786e\u800c\u5bfc\u81f4\u8fc7\u591a\u4f7f\u7528 GPU#0\uff09\uff0c\u6211\u4eec\u89e3\u51b3\u4e86\u8fd9\u4e9b\u95ee\u9898\u3002<\/p>\n<p>3.CPU\/RAM \u5185\u5b58\u4e0d\u8db3\uff08OOM\uff09\u9519\u8bef\uff0c\u8fd9\u4e9b\u9519\u8bef\u5728\u9519\u8bef\u65e5\u5fd7\u4e2d\u4e0d\u592a\u5bb9\u6613\u53d1\u73b0\uff0c\u5e76\u4e14\u901a\u5e38\u80fd\u901a\u8fc7 Docker \u5bb9\u5668\u5916\u7684\u4e3b\u673a\u7684 dmesg \u65e5\u5fd7\u68c0\u6d4b\u51fa\u6765\u3002\u5f53 OOM Killer \u8c03\u7528\u505c\u6b62\u4e00\u4e2a\u5206\u53c9\u8fdb\u7a0b\u6216\u540c\u7ea7\u7f51\u7edc\uff08network peer\uff09\u65f6\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u5b83\u4eec\u4e3b\u8981\u8868\u73b0\u4e3a CalledProcessError \u6216 ConnectionError\u3002\u5f53\u4ece dmesg \u68c0\u6d4b\u5230\u4e86 OOM Killer \u8c03\u7528\u65f6\uff0c\u6211\u4eec\u66f4\u503e\u5411\u4e8e\u76f4\u63a5\u653e\u5f03\u5065\u5eb7\u68c0\u67e5\uff0c\u5e76\u91cd\u542f\u8be5\u673a\u7bb1\u3002\u6211\u4eec\u8fd8\u68c0\u67e5\u4e86\u6211\u4eec\u7684\u4ee3\u7801\u8def\u5f84\u662f\u5426\u6709\u8db3\u591f\u7684\u624b\u52a8\u5783\u573e\u6536\u96c6\uff08\u4e0b\u9762\u6709\u4e00\u90e8\u5206\u4ecb\u7ecd\u4e86\u5982\u4f55\u7981\u7528\u5b83\uff09\uff0c\u5e76\u8fd8\u68c0\u67e5\u4e86\u662f\u5426\u6709\u610f\u5916\u5c1d\u8bd5\u8fdb\u884c\u8ba1\u7b97\u6216\u5c06<mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u79fb\u52a8\u5230 CPU \u4e0a\u3002<\/p>\n<ul>\n<li>\n<p>\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5d29\u6e83<\/p>\n<\/li>\n<\/ul>\n<p>\u9996\u8981\u4efb\u52a1\u662f\u8ba9\u7cfb\u7edf\u80fd\u81ea\u52a8\u8fd0\u884c\uff0c\u8ba9\u5176\u80fd\u81ea\u52a8\u91cd\u65b0\u8fd0\u884c\u6240\u6709\u5065\u5eb7\u68c0\u67e5\uff0c\u7136\u540e\u5728\u6ca1\u53d1\u73b0\u4e0d\u5065\u5eb7\u4e3b\u673a\u65f6\u91cd\u542f\u8fd0\u884c\u3002\u6211\u4eec\u9047\u5230\u4e86\u4e00\u4e9b\u968f\u673a\u7684\u786c\u4ef6\u9519\u8bef\uff0c\u5305\u62ec Xid \u548c SXid \u9519\u8bef\uff1b\u8fd9\u4e9b\u9519\u8bef\u53ef\u80fd\u4f1a\u5bfc\u81f4\u8fd0\u884c\u5d29\u6e83\uff0c\u5374\u4e0d\u4f1a\u53d1\u51fa\u6709\u610f\u4e49\u7684 Python \u6808\u8ddf\u8e2a\u8bb0\u5f55\u3002\u884c\u91cd<mark data-type=\"concepts\" data-id=\"8ec6a68f-ad96-4b85-ab72-6f8931886922\">\u6620\u5c04<\/mark>\u7b49\u4e00\u4e9b\u95ee\u9898\u53ef\u901a\u8fc7\u91cd\u542f\u6062\u590d\u3002\u4e0d\u53ef\u7ea0\u6b63\u7684 ECC \u9519\u8bef\u7b49\u53e6\u4e00\u4e9b\u95ee\u9898\u5219\u5f80\u5f80\u9700\u8981\u786c\u4ef6\u7ef4\u62a4\u6216\u66f4\u6362\u96f6\u4ef6\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u89c2\u5bdf\u5230\u683c\u5f0f\u9519\u8bef\u7684\u8bad\u7ec3\u6570\u636e\u4e5f\u4f1a\u5bfc\u81f4\u5d29\u6e83\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5982\u679c<mark data-type=\"concepts\" data-id=\"930c591c-a35b-4761-83ef-22ef12aa3c5f\">\u8bed\u6599\u5e93<\/mark>\u4e2d\u6709\u4e00\u4e2a\u975e\u5e38\u5927\u7684\u5355\u4e2a\u6587\u6863\uff0c\u5c31\u53ef\u80fd\u5bfc\u81f4 GPU \u6216 CPU \u51fa\u73b0\u5185\u5b58\u4e0d\u8db3\u9519\u8bef\u3002\u4e3a\u4e86\u9632\u6b62\u51fa\u73b0\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u5b8c\u5168\u786e\u5b9a\u5f0f\u7684\u6570\u636e\u52a0\u8f7d\u5668 \u2014\u2014 \u901a\u8fc7\u4e0e epoch \u6216\u6b65\u6570\u76f8\u5173\u8054\uff0c\u8ba9\u6bcf\u4e00\u6b21\u5d29\u6e83\u90fd\u53ef\u8f7b\u677e\u590d\u73b0\u3002\u6211\u4eec\u53d1\u73b0\u7981\u7528\u6570\u636e\u52a0\u8f7d\u6216\u66ff\u6362\u5047\u6570\u636e\uff08\u4f8b\u5982\u5168\u96f6\u6570\u636e\uff09\u6709\u52a9\u4e8e\u786e\u8ba4\u9519\u8bef\u7684\u6839\u672c\u539f\u56e0\u662f\u5426\u662f\u6570\u636e\u3002<\/p>\n<p>\u6700\u540e\uff0c\u901a\u8fc7\u6307\u6807\u805a\u5408\u65b9\u6cd5\u8bb0\u5f55\u7f51\u7edc\u548c\u4e00\u822c\u8282\u70b9\u7684\u5065\u5eb7\u7edf\u8ba1\u6570\u636e\u4e5f\u5f88\u6709\u5e2e\u52a9\u3002\u4ee5\u592a\u7f51\u77ed\u6682\u65ad\u5f00\u6216\u78c1\u76d8\u7a7a\u95f4\u4e0d\u8db3\u7b49\u95ee\u9898\u53ef\u80fd\u4e0d\u4f1a\u663e\u793a\u4e3a\u6709\u7528\u7684\u9519\u8bef\u6d88\u606f\uff0c\u4f46\u5374\u53ef\u4ee5\u5f88\u8f7b\u677e\u5730\u4e0e\u5df2\u6536\u96c6\u7684\u6570\u636e\u76f8\u5173\u8054\u3002<\/p>\n<ul>\n<li>\n<p>\u6ca1\u6709\u6808\u8ddf\u8e2a\u4fe1\u606f\u7684\u6302\u8d77\uff08\u4e4b\u540e\u53ef\u80fd\u4f1a\u6709\u8d85\u65f6\u95ee\u9898\uff09<\/p>\n<\/li>\n<\/ul>\n<p>\u7531\u4e8e\u8fd9\u4e9b\u95ee\u9898\u7f3a\u4e4f\u6709\u5e2e\u52a9\u7684\u4fe1\u606f\uff0c\u52a0\u4e0a\u5f88\u96be\u53ef\u9760\u5730\u590d\u73b0\uff0c\u56e0\u6b64\u8fd9\u7c7b\u9519\u8bef\u7684\u8c03\u8bd5\u5de5\u4f5c\u7740\u5b9e\u8ba9\u4eba\u6cae\u4e27\u3002<\/p>\n<p>\u5176\u4e2d\u6700\u4ee4\u4eba\u96be\u5fd8\u7684\u9519\u8bef\u7c7b\u578b\u4f34\u968f\u7740\u8fd9\u6837\u7684\u62a5\u9519\u4fe1\u606f\uff1a<\/p>\n<section>\n<pre data-lang=\"cs\"><code>Watchdog caught collective operation timeout:WorkNCCL (SeqNum=408951, OpType=_ALLGATHER_BASE, \u2026 , Timeout (ms)=600000) ran for 600351 milliseconds before timing out<\/code><\/pre>\n<\/section>\n<p>\u5e76\u4e14\u8bad\u7ec3\u8fd0\u884c\u4e2d\u7684\u6240\u6709 GPU \u5de5\u4f5c\u5668\u90fd\u53d1\u51fa\u4e86\u8fd9\u6837\u7684\u62a5\u9519\u4fe1\u606f\u3002<\/p>\n<p>\u8fd9\u610f\u5473\u7740\u4e00\u53f0\u6216\u591a\u53f0\u4e3b\u673a\u672a\u80fd\u5b8c\u6210 NCCL \u64cd\u4f5c\u6216\u8005 NCCL \u548c InfiniBand \u8fde\u63a5\u5d29\u6e83\uff0c\u5bfc\u81f4\u5176\u4ed6\u6240\u6709\u4e3b\u673a\u540c\u65f6\u5361\u5728\u4e86\u67d0\u4e2a<mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u8fd0\u7b97\u4e0a\uff0c\u76f4\u5230\u8fbe\u5230 NCCL_TIMEOUT \u8d85\u65f6\u65f6\u95f4\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u53d7 NCCL \u8f6f\u4ef6\u5e93\u672c\u8d28\u6240\u9650\uff0c\u6211\u4eec\u5f88\u96be\u627e\u5230\u7a76\u7adf\u662f\u54ea\u53f0\u4e3b\u673a\u51fa\u4e86\u95ee\u9898\u3002<\/p>\n<p>\u6211\u4eec\u5bf9 NCCL \u8f6f\u4ef6\u5e93\u7684\u65e5\u5fd7\u8bb0\u5f55\u505a\u4e86\u4e00\u4e9b\u4fee\u6539\uff0c\u53c2\u89c1\u6211\u4eec\u7684\u5206\u53c9\u7248\uff1ahttps:\/\/github.com\/boweiliu\/nccl \u3002\u4ece\u800c\u5728\u5d29\u6e83\u53d1\u751f\u65f6\u53ef\u80fd\u66f4\u597d\u5730\u63ed\u793a\u6b63\u5728\u6267\u884c\u7684\u6d88\u606f\u6216\u64cd\u4f5c\uff0c\u4ece\u800c\u786e\u5b9a\u963b\u6b62\u8fd0\u884c\u7684\u53ef\u80fd\u662f\u54ea\u53f0\u4e3b\u673a\u6216 GPU\u3002<\/p>\n<p>\u8bf7\u6ce8\u610f\uff0c\u4e3a\u4e86\u8bc6\u522b\u884c\u4e3a\u4e0d\u5f53\u7684\u4e3b\u673a\uff0c\u6211\u4eec\u901a\u5e38\u9700\u8981\u641e\u6e05\u695a\u54ea\u4e9b\u4e3b\u673a\u6ca1\u6709\u751f\u6210\u67d0\u4e9b\u65e5\u5fd7\u6d88\u606f\u3002\u7f3a\u5c11\u6b64\u7c7b\u6d88\u606f\u8868\u660e\u8be5\u4e3b\u673a\u4e0a\u7684\u5de5\u4f5c\u5668\u5df2\u843d\u540e\u6216\u5df2\u5d29\u6e83\u3002<\/p>\n<p>\u5176\u4ed6\u6ca1\u6709\u53ef\u7528\u9519\u8bef\u6d88\u606f\u7684\u65e0\u54cd\u5e94\u60c5\u51b5\u901a\u5e38\u4e0e\u786c\u4ef6\u76f8\u5173\u95ee\u9898\u6709\u5173\uff0c\u6bd4\u5982\u4e4b\u524d\u63d0\u5230\u7684 Xid\/SXid\/ECC \u9519\u8bef\u4f1a\u5bfc\u81f4\u82f1\u4f1f\u8fbe\u9a71\u52a8\u6216\u82f1\u4f1f\u8fbe Docker \u901a\u4fe1\u9a71\u52a8\u9501\u5b9a\u3002\u4e3a\u4e86\u533a\u5206 NCCL \u6302\u8d77\u4e0e\u9a71\u52a8\u6302\u8d77\u4ee5\u53ca Python \u4ee3\u7801\u4e2d\u7684\u7ade\u4e89\u6761\u4ef6\u6216\u6b7b\u9501\uff0c\u6211\u4eec\u4f7f\u7528 Py-Spy \u548c GNU Project Debugger\uff08GDB\uff09\u7b49\u5de5\u5177\u6765\u5b9e\u65f6\u8c03\u8bd5\u9047\u5230\u7684\u505c\u6ede\u8fdb\u7a0b\u3002\u4f7f\u7528\u6b64\u65b9\u6cd5\u53ef\u53d1\u73b0\u4e00\u4e2a\u7279\u5b9a\u95ee\u9898\uff1a\u7531\u4e8e Python \u7ebf\u7a0b\u8bbe\u7f6e\u4e2d\u7684\u914d\u7f6e\u9519\u8bef\uff0c\u6211\u4eec\u65e0\u6cd5\u5728\u67d0\u4e9b\u4e3b\u673a\u4e0a\u6b63\u786e\u542f\u52a8\u516b\u4e2a\u591a\u7ebf\u7a0b NCCL GPU \u8fdb\u7a0b\uff0c\u8fd9\u4e9b\u8fdb\u7a0b\u5728 PyTorch \u4e4b\u524d\u7684\u521d\u59cb\u5316\u4ee3\u7801\u9636\u6bb5\u9047\u5230\u4e86\u7ade\u4e89\u6761\u4ef6\u3002<\/p>\n<ul>\n<li>\n<p>\u8bad\u7ec3\u51cf\u901f\uff08\u7531 MFU \u6d4b\u91cf\uff09<\/p>\n<\/li>\n<\/ul>\n<p>\u7f3a\u4e4f\u5de5\u5177\u8ba9\u8fd9\u7c7b\u95ee\u9898\u6bd4\u524d\u4e00\u7c7b\u66f4\u8ba9\u4eba\u6cae\u4e27\u3002\u9664\u4e86\u4f7f\u7528 Py-Spy\u3001\u6808\u8ddf\u8e2a\u68c0\u67e5\u548c GDB \u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u91c7\u7528\u4e86 NVIDIA Nsight \u548c profiling \u5de5\u5177\uff0c\u5176\u4e2d\u4e00\u4e9b\u5de5\u5177\u5728\u9ad8\u5ea6\u5206\u5e03\u5f0f\u7684\u8bbe\u7f6e\u4e2d\u5f88\u96be\u4f7f\u7528\u3002<\/p>\n<p>\u9057\u61be\u7684\u662f\uff0c\u5bfc\u81f4\u4e00\u822c\u51cf\u901f\u6216\u8ba9\u901f\u5ea6\u4f4e\u4e8e\u4e4b\u524d\u6f14\u793a\u7684\u6a21\u578b\u6d6e\u70b9\u6570\u5229\u7528\u7387\uff08MFU\uff09\u7684\u539f\u56e0\u6709\u5f88\u591a\u3002<\/p>\n<p>\u9996\u5148\uff0c\u4e8b\u5b9e\u8bc1\u660e\u591a\u6b21\u68c0\u67e5\u914d\u7f6e\u3001\u4ee3\u7801\u548c\u73af\u5883\u53d8\u91cf\u662f\u6709\u7528\u7684\u3002\u6211\u4eec\u7ecf\u5386\u8fc7\u7684\u9519\u8bef\u5305\u62ec\uff1a\u8fd0\u884c\u4e86\u9519\u8bef\u7684\u6a21\u578b\u3001\u6279\u91cf\u5927\u5c0f\u9519\u8bef\u3001UFM \u6216 NCCL \u8bbe\u7f6e\u51fa\u9519\u3001CUDA_DEVICE_MAX_CONNECTIONS \u51fa\u9519\u3002\u8fd9\u90fd\u4f1a\u5bfc\u81f4\u6027\u80fd\u65e0\u6cd5\u8fbe\u5230\u6700\u4f18\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u53d1\u73b0\u6d4b\u91cf\u77ac\u65f6\uff08\u5373\u6bcf<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\uff09MFU\uff08\u800c\u4e0d\u662f\u5e73\u6ed1\u6216\u7a97\u53e3\u5e73\u5747\u503c\uff09\u5f88\u6709\u7528\uff0c\u56e0\u4e3a\u672a\u5e73\u6ed1\u5904\u7406\u7684 MFU \u66f2\u7ebf\u901a\u5e38\u6709\u52a9\u4e8e\u8bca\u65ad\u95ee\u9898\u7c7b\u522b\u3002\u5bfc\u81f4\u8bad\u7ec3\u901f\u5ea6\u4e0b\u964d\u7684\u95ee\u9898\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p>\u4ece\u975e\u5e38\u4f4e\u7684 MFU\uff08\u4f4e\u4e8e\u9884\u671f\u7684\u5341\u5206\u4e4b\u4e00\uff09\u7acb\u5373\u5f00\u59cb\u8bad\u7ec3\u5e76\u4fdd\u6301\u7a33\u5b9a<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u591a\u534a\u662f InfiniBand \u7f51\u7edc\u8fde\u63a5\u7684\u786c\u4ef6\u95ee\u9898\uff0c\u4f8b\u5982 T2 \u6216 T3 \u5c42\u7684\u4ea4\u6362\u673a\u6b7b\u673a\u3002GPU \u548c NIC \u4e4b\u95f4\u7684\u786c\u4ef6\u95ee\u9898\u4e5f\u53ef\u80fd\u5bfc\u81f4\u8be5\u60c5\u51b5\uff0c\u5bf9\u6b64 dmesg \u4f1a\u8fd9\u6837\u62a5\u9519\uff1aPCIe x16 lanes limited by \u2026<\/p>\n<ul>\n<li>\n<p>\u4ece\u9884\u671f MFU \u7684 30% \u7acb\u5373\u5f00\u59cb\u8bad\u7ec3\u5e76\u4fdd\u6301\u7a33\u5b9a<\/p>\n<\/li>\n<\/ul>\n<p>\u5176\u539f\u56e0\u53ef\u80fd\u662f\u4e00\u53f0\u4e3b\u673a\u7684 GDR \u8bbe\u7f6e\u4e0d\u6b63\u786e\uff08NVIDIA \u5bf9\u7b49\u5185\u5b58\uff09\u6216 GDR \u73af\u5883\u53d8\u91cf\u4e0d\u6b63\u786e\u3002<\/p>\n<ul>\n<li>\n<p>\u4ece\u9884\u671f MFU \u7684\u7ea6 60-80% \u7acb\u5373\u5f00\u59cb\u8bad\u7ec3\u5e76\u4fdd\u6301\u7a33\u5b9a<\/p>\n<\/li>\n<\/ul>\n<p>\u6700\u5e38\u89c1\u7684\u539f\u56e0\u662f InfiniBand \u94fe\u8def\u8d28\u91cf\u4e0d\u884c\u6216\u6545\u969c\uff0c\u5c24\u5176\u662f\u5355\u53f0 GPU \u51fa\u73b0\u4e86\u4e0e InfiniBand NIC \u76f8\u5173\u7684\u6545\u969c\uff0c\u5bfc\u81f4 NCCL \u5c1d\u8bd5\u7ecf\u7531\u672c\u5730 NVLink \u8def\u7531\u6d41\u91cf\u5e76\u5728\u540c\u4e00\u4e3b\u673a\u4e0a\u7684\u53e6\u4e00\u53f0 GPU \u4e0a\u4f7f\u7528 NIC\u3002CPU \u8282\u6d41\u4e5f\u53ef\u80fd\u5bfc\u81f4\u8be5\u95ee\u9898\uff0c\u8fd9\u9700\u8981\u8c03\u6574\u67d0\u4e9b\u4e3b\u673a\u7684 BIOS \u8bbe\u7f6e\u3002<\/p>\n<ul>\n<li>\n<p>\u5728\u5904\u7406\u67d0\u4e9b\u6570\u636e<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u65f6\u7a81\u7136\u5927\u5e45\u51cf\u901f\uff08\u4e0b\u964d 10 \u500d\uff09\uff0c\u5e76\u4e14\u7ecf\u5e38\u53d1\u751f\u8fd9\u79cd\u60c5\u51b5<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u57fa\u672c\u4e0a\u90fd\u4e0e\u68c0\u67e5\u70b9\u6216\u8bc4\u4f30\u6709\u5173 \u2014\u2014 \u53ef\u901a\u8fc7\u68c0\u67e5 epoch \u6570\u6216\u6b65\u6570\u6765\u9a8c\u8bc1\u3002\u607c\u706b\u7684\u662f\uff0c\u5982\u679c\u8bbe\u7f6e\u4e86\u5728 MFU \u5f02\u5e38\u65f6\u81ea\u52a8\u544a\u8b66\uff0c\u5c31\u4f1a\u51fa\u73b0\u8bb8\u591a\u8bef\u62a5\u3002<\/p>\n<ul>\n<li>\n<p>\u5728\u5904\u7406\u67d0\u4e9b\u6570\u636e<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u65f6\u7a81\u7136\u5927\u5e45\u51cf\u901f\uff08\u4e0b\u964d 10 \u500d\uff09<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u79cd\u60c5\u51b5\u662f\u968f\u673a\u53d1\u751f\u7684\u5e76\u4e14\u76f8\u5f53\u7f55\u89c1\uff08\u5927\u6982\u6bcf 15 \u5206\u949f\u4e00\u6b21\uff09\uff0c\u5e76\u4e14\u4e4b\u540e\u9a6c\u4e0a\u5c31\u4f1a\u5b8c\u5168\u6062\u590d\u5230\u826f\u597d\u7684 MFU\u3002<\/p>\n<p>\u6700\u5e38\u89c1\u7684\u539f\u56e0\u4f3c\u4e4e\u662f\u5176\u4ed6\u9700\u8981\u5927\u91cf CPU \u8ba1\u7b97\u7684\u5de5\u4f5c\u8d1f\u8f7d\u88ab<mark data-type=\"tech_methods\" data-id=\"1ef8dd91-7c80-46f2-811e-0722d3d333d4\">\u8c03\u5ea6<\/mark>\u5230\u4e86\u4e00\u53f0\u8fd0\u884c\u4e2d\u7684\u4e3b\u673a\u4e0a\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u4e0e\u5176\u6784\u5efa\u5206\u6790\u5de5\u5177\u6765\u8bc6\u522b\u7279\u5b9a\u7684\u4e3b\u673a\uff0c\u901a\u8fc7 PID \u6765\u7c97\u7565\u5730\u76d1\u63a7 CPU \u4f1a\u66f4\u5bb9\u6613\u3002\u5176\u539f\u56e0\u53ef\u80fd\u662f\u5076\u5c14\u51fa\u73b0\u7684\u7f51\u7edc\u8fde\u63a5\u95ee\u9898\uff0c\u6bd4\u5982\u6570\u636e\u52a0\u8f7d\u5668\u906d\u9047\u74f6\u9888\u3002\u6211\u4eec\u76d1\u63a7\u4e86\u6570\u636e\u52a0\u8f7d\u3001\u68c0\u67e5\u70b9\u548c\u4efb\u4f55\u975e NCCL \u4ee3\u7801\u7684\u6307\u6807\u6570\u636e\u5e76\u6dfb\u52a0\u4e86 Python \u4ee3\u7801\u8ba1\u65f6\u65e5\u5fd7\uff0c\u4e8b\u5b9e\u8bc1\u660e\u8fd9\u975e\u5e38\u53ef\u9760\u3002<\/p>\n<ul>\n<li>\n<p>MFU \u5728\u8fd0\u884c\u8fc7\u7a0b\u4e2d\u9010\u6e10\u51cf\u6162\uff0c\u4f46\u6bcf\u6b21\u91cd\u542f\u540e\u53c8\u4f1a\u56de\u5230 100%<\/p>\n<\/li>\n<\/ul>\n<p>\u7406\u8bba\u4e0a\u8bb2\uff0c\u5176\u539f\u56e0\u53ef\u80fd\u662f\u4ea4\u6362\u673a\u4e0a\u7684\u70ed\u91cf\u79ef\u7d2f\uff0c\u4f46\u6211\u4eec\u4ece\u672a\u89c1\u8fc7\u8fd9\u79cd\u60c5\u51b5\u3002\u4e0d\u8fc7\uff0c\u6211\u4eec\u4f7f\u7528 Python \u548c NVIDIA \u5206\u6790\u5668\u786e\u5b9a\uff1a\u6027\u80fd\u4e0b\u964d\u7684\u539f\u56e0\u4f3c\u4e4e\u662f\u81ea\u52a8\u5783\u573e\u6536\u96c6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b5352643a56db241fe9bb8525aa4cbf1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b5352643a56db241fe9bb8525aa4cbf1.png\"><\/a><\/p>\n<p>\u5728\u8c03\u8bd5\u89e3\u51b3\u8fd9\u4e9b\u51cf\u901f\u95ee\u9898\u65f6\uff0c\u6211\u4eec\u53d1\u73b0\u541e\u5410\u91cf\u51e0\u4e4e\u5fc5\u7136\u4f1a\u5468\u671f\u6027\u5730\u4e0b\u964d\u3002\u968f\u7740\u8bad\u7ec3\u5730\u63a8\u8fdb\uff0c\u8fd9\u79cd\u4e0b\u964d\u4f1a\u5bf9\u5206\u5e03\u5f0f\u8fd0\u7b97\u5e26\u6765\u8d8a\u6765\u8d8a\u591a\u7684\u5f71\u54cd\u3002\u8fd9\u8ba9\u6211\u4eec\u731c\u60f3\u4e0b\u964d\u7684\u539f\u56e0\u53ef\u80fd\u4e0e\u81ea\u52a8\u5783\u573e\u6536\u96c6\u6709\u5173 \u2014\u2014 \u6211\u4eec\u901a\u8fc7\u5206\u6790\u548c\u6d4b\u8bd5\u9a8c\u8bc1\u4e86\u8fd9\u4e2a\u731c\u60f3\u3002\u5f53\u6211\u4eec\u7981\u7528\u4e86\u81ea\u52a8\u5783\u573e\u6536\u96c6\uff0c\u5e76\u5728\u6240\u6709\u4e3b\u673a\u4e0a\u8bbe\u5b9a\u53ea\u5728\u7279\u5b9a\u7684\u95f4\u9694\u5185\u6536\u96c6\u5783\u573e\uff0c\u8fd9\u79cd\u541e\u5410\u91cf\u300c\u4e0b\u964d\u300d\u5c31\u6d88\u5931\u4e86\u3002<\/p>\n<p>\u6211\u4eec\u4f7f\u7528\u4e86\u4e00\u79cd\u57fa\u4e8e ZeRO-3 \u7684\u540c\u6b65\u5206\u5e03\u5f0f\u8bad\u7ec3\u7b97\u6cd5 FSDP\u3002\u5728\u963b\u585e\u64cd\u4f5c\u671f\u95f4\uff0c\u8fd0\u884c\u5783\u573e\u6536\u96c6\u7684\u5355\u4e2a\u5de5\u4f5c\u5668\u8fdb\u7a0b\u53ef\u80fd\u4f1a\u51cf\u6162\u5176\u4ed6\u6240\u6709\u5de5\u4f5c\u5668\u7684\u901f\u5ea6\u3002\u5982\u679c\u6709\u6570\u767e\u4e2a\u5de5\u4f5c\u5668\u8fdb\u7a0b\uff0c\u5c31\u53ef\u80fd\u5bfc\u81f4\u901f\u5ea6\u5927\u5e45\u4e0b\u964d\u3002<\/p>\n<p>\u4e00\u5f00\u59cb\u6027\u80fd\u826f\u597d\uff0c\u7136\u540e\u7a81\u7136\u4e0b\u964d\uff08\u8fbe\u5230\u9884\u671f\u7684 70%\uff09\uff0c\u5e76\u4e14\u4ee5\u9ad8\u9891\u6301\u7eed\uff08\u6bcf 15 \u79d2\uff09<\/p>\n<p>\u6211\u4eec\u89c2\u5bdf\u5230\u8fd9\u4e0e\u82f1\u4f1f\u8fbe GPU \u7684\u300c\u65f6\u949f\u8282\u6d41\u539f\u56e0\u300d\u76f8\u5173\uff0c\u8fd9\u53ef\u901a\u8fc7\u5bf9\u82f1\u4f1f\u8fbe DCGM \u8fdb\u884c\u9002\u5f53\u7684\u8bbe\u7f6e\u6765\u89e3\u51b3\u3002\u53d1\u70ed\u95ee\u9898\uff08GPU \u9ad8\u6e29\u6216\u4e3b\u673a\u51b7\u5374\u98ce\u6247\u6545\u969c \/ \u6548\u679c\u4e0b\u964d\uff09\u6216\u7535\u6e90\u6545\u969c\u4f1a\u5bfc\u81f4\u8be5\u95ee\u9898\u3002\u53e6\u5916\uff0c\u5f53\u6211\u4eec\u540c\u65f6\u6700\u5927\u5316\u6240\u6709 8 \u53f0 GPU \u5229\u7528\u7387\u548c 8x NIC InfiniBand \u5229\u7528\u7387\u4ee5\u53ca CPU\/RAM\/ \u78c1\u76d8\u65f6\uff0c\u6211\u4eec\u7684\u4e00\u4e9b\u5177\u6709\u7279\u5b9a\u7535\u6e90\u786c\u4ef6\u7684\u4e3b\u673a\u4f1a\u51fa\u73b0\u7535\u538b\u95ee\u9898\uff0c\u4f46\u53ea\u6709\u5168\u90e8\u4f7f\u7528\u5b83\u4eec\uff08\u901a\u5e38\u53ea\u5728\u5b9e\u9645\u8bad\u7ec3\u8fd0\u884c\u671f\u95f4\uff09\u65f6\u624d\u4f1a\u51fa\u73b0\u8fd9\u79cd\u60c5\u51b5\u3002<\/p>\n<ul>\n<li>\n<p>\u6027\u80fd\u4f18\u826f\u4f46\u566a\u58f0\u6bd4\u5e73\u5e38\u60c5\u51b5\u591a\uff08\u9ad8\u9891\u767d\u566a\u58f0\u65b9\u5dee\u5728\u9884\u671f MFU \u7684 90% \u548c 100% \u4e4b\u95f4\uff09<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u4e5f\u4e0e InfiniBand \u786c\u4ef6\u6709\u5173\uff0c\u4f46\u901a\u5e38\u662f\u7531\u4e8e\u7f51\u7edc\u4e2d\u8f83\u9ad8\u5c42\u7684\u94fe\u8def\u51fa\u73b0\u4e00\u5b9a\u7a0b\u5ea6\u7684\u6027\u80fd\u4e0b\u964d\u6216\u6296\u52a8\uff0c\u800c\u4e0d\u662f\u5197\u4f59\u5ea6\u8f83\u4f4e\u7684\u4e3b\u673a\u5230 T2 \u5c42\u3002<\/p>\n<p>\u4e0d\u5e78\u7684\u662f\uff0c\u5f88\u591a\u8fd9\u7c7b\u95ee\u9898\u90fd\u96be\u4ee5\u5b9a\u4f4d\u5230\u67d0\u53f0\u5177\u4f53\u4e3b\u673a\uff0c\u800c\u4e0e InfiniBand \u76f8\u5173\u7684\u95ee\u9898\u5c24\u5176\u96be\u4ee5\u5b9a\u4f4d\uff0c\u8fd9\u662f\u7531\u4e8e InfiniBand \u4ea4\u6362\u673a\u6280\u672f\u7684\u62d3\u6251<mark data-type=\"concepts\" data-id=\"1ccbfcc7-1b56-44d6-b64b-729d855abcb1\">\u611f\u77e5<\/mark>\u7279\u6027\u3002InfiniBand \u4f3c\u4e4e\u66f4\u504f\u597d InfiniBand fat-tree \u8bbe\u8ba1\u4e2d\u7684\u76f8\u90bb\u4e3b\u673a\uff0c\u800c UFM \u80fd\u4ee5\u4e0d\u5bf9\u79f0\u7684\u94fe\u8def\u901f\u5ea6\u8def\u7531\u6570\u636e\u5305\u3002<\/p>\n<p>\u4ee5\u4e0b\u662f\u7528\u4e8e\u8c03\u8bd5\u541e\u5410\u91cf\u95ee\u9898\u7684\u7b80\u5355\u6458\u8981 \/ \u6d41\u7a0b\u56fe \/ \u5b8c\u5907\u6027\u68c0\u67e5\u8868\uff1a<\/p>\n<ul>\n<li>\n<p>\u8fd9\u5957\u7cfb\u7edf\u4e4b\u524d\u80fd\u6b63\u5e38\u5de5\u4f5c\u5417\uff1f<\/p>\n<\/li>\n<li>\n<p>\u4f60\u6700\u8fd1\u505a\u4e86\u4ec0\u4e48\u4fee\u6539\uff08\u6bd4\u5982\u5408\u5e76\u4ee3\u7801\u3001\u66f4\u65b0\u9a71\u52a8\uff09\uff1f<\/p>\n<\/li>\n<li>\n<p>\u4f60\u8fd0\u884c\u7684\u4e3b\u673a\u662f\u5426\u5065\u5eb7\uff1f\u4f60\u7684\u6240\u6709\u4f9d\u8d56\u670d\u52a1\u662f\u5426\u90fd\u8fd0\u884c\u6b63\u5e38\uff0c\u5305\u62ec\u7b2c\u4e09\u65b9\u7684 SaaS\uff0c\u6bd4\u5982\u00a0Docker Hub\u3001GitHub \u7b49\u7b49\uff1f<\/p>\n<\/li>\n<li>\n<p>\u4f60\u80fd\u786e\u5b9a\u73b0\u5728\u8fd0\u884c\u7684\u4ee3\u7801\u3001\u73af\u5883\u3001\u914d\u7f6e\u3001\u7248\u672c\u3001\u4e3b\u673a\u5217\u8868\u3001\u6392\u540d\u987a\u5e8f\u3001\u968f\u673a\u79cd\u5b50\u4e0e\u4e0a\u6b21\u5b8c\u5168\u76f8\u540c\u5417\uff1f\uff08\u5982\u679c\u80fd\u5b9e\u73b0\u8fd9\u6837\u7684\u68c0\u67e5\u7684\u8bdd\u3002\uff09<\/p>\n<\/li>\n<li>\n<p>\u95ee\u9898\u53ef\u590d\u73b0\u5417\uff1f<\/p>\n<\/li>\n<li>\n<p>\u4e0e\u5176\u4ed6\u4e8b\u7269\u6709\u4f55\u5173\u8054\uff1f\u5176\u4ed6\u8fdb\u7a0b\uff1f\u6bcf\u65e5 crontab \u5b9a\u65f6\u4efb\u52a1\uff1f\u4e3b\u673a\u6216 DCGM \u6216 UFM \u6307\u6807\uff1f<\/p>\n<\/li>\n<li>\n<p>\u4f60\u7684\u5de5\u5177\u662f\u5426\u80fd\u6b63\u786e\u5ea6\u91cf\u8fd9\u4e9b\u6307\u6807\uff1f<\/p>\n<\/li>\n<li>\n<p>\u5728\u8fd0\u884c\u7ea6\u7b80\u7248\u7684\u4ee3\u7801\uff08\u4f7f\u7528\u66f4\u5c0f\u7684\u6a21\u578b\u3001\u5047\u6570\u636e\u3001\u4e0d\u4fdd\u5b58\u6216\u52a0\u8f7d\u68c0\u67e5\u70b9\uff09\u65f6\uff0c\u95ee\u9898\u662f\u5426\u4f9d\u7136\u5b58\u5728\uff1f<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u6539\u8fdb\u57fa\u7840\u8bbe\u65bd\u5de5\u5177<\/strong><\/p>\n<p>\u5b8c\u6210\u4e86\u4ee5\u4e0a\u6b65\u9aa4\u4e4b\u540e\uff0c\u5c31\u80fd\u5728\u8bad\u7ec3\u6a21\u578b\u65f6\u5b9e\u73b0\u4f18\u826f\u6027\u80fd\u4e86\u2026\u2026 \u81f3\u5c11\u5728\u67d0\u4e2a\u5730\u65b9\u51fa\u6545\u969c\u4e4b\u524d\u662f\u8fd9\u6837\u3002<\/p>\n<p>\u672c\u8282\u5c06\u4ecb\u7ecd\u4e00\u4e9b\u7528\u4e8e\u786e\u4fdd\u8bad\u7ec3\u6301\u7eed\u7a33\u5b9a\u7684\u5de5\u5177\u548c\u7cfb\u7edf\uff0c\u540c\u65f6\u6700\u597d\u5c3d\u53ef\u80fd\u5730\u5c11\u5730\u9700\u8981\u4eba\u7c7b\u5e72\u9884\u3002\u7531\u4e8e\u6211\u4eec\u8fd9\u4e2a\u56e2\u961f\u5f88\u5c0f\uff0c\u56e0\u6b64\u6211\u4eec\u6ca1\u6709\u8db3\u591f\u7684\u4eba\u624b\u6765\u8fdb\u884c\u4eba\u5de5\u7ef4\u4fee\uff0c\u6240\u4ee5\u6211\u4eec\u4e5f\u5e0c\u671b\u80fd\u5c3d\u53ef\u80fd\u5730\u81ea\u52a8\u5316\u8fd9\u4e2a\u8fc7\u7a0b\u3002<\/p>\n<p>\u6211\u4eec\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u9047\u5230\u7684\u6240\u6709\u95ee\u9898\u51e0\u4e4e\u90fd\u53ef\u5f52\u56e0\u4e8e\u673a\u5668\u6216\u7f51\u7edc\u7ec4\u4ef6\u6545\u969c\u3002\u8fd9\u7c7b\u6545\u969c\u5728\u5927\u578b\u96c6\u7fa4\u4e2d\u5f88\u5e38\u89c1\uff0c\u56e0\u6b64\u6211\u4eec\u7684\u505a\u6cd5\u662f\u81ea\u52a8\u7981\u7528\u51fa\u6545\u969c\u7684\u673a\u5668\u548c\u7f51\u7edc\u7ec4\u4ef6\u5e76\u53d1\u9001\u7ef4\u4fee\u8bf7\u6c42\u3002<\/p>\n<ul>\n<li>\n<p>\u673a\u5668\u6545\u969c<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u7cfb\u7edf\uff0c\u53ef\u5728\u8fd0\u884c\u5d29\u6e83\u65f6\u81ea\u52a8\u4ece\u6700\u8fd1\u7684\u68c0\u67e5\u70b9\u91cd\u542f\u3002\u5728\u8fd9\u4e2a\u91cd\u542f\u8fc7\u7a0b\u4e2d\uff0c\u9996\u5148\u662f\u5bf9\u6bcf\u53f0\u53ef\u7528\u673a\u5668\u8fdb\u884c\u5065\u5eb7\u68c0\u67e5\uff0c\u7136\u540e\u57fa\u4e8e\u5176\u4f20\u9012\u7684\u5065\u5eb7\u68c0\u67e5\u7ed3\u679c\u5bf9\u6bcf\u53f0\u673a\u5668\u8fdb\u884c\u5206\u7c7b\uff1b\u7136\u540e\u5c1d\u8bd5\u5728\u6700\u5065\u5eb7\u7684\u673a\u5668\u4e0a\u91cd\u542f\u8bad\u7ec3\u3002<\/p>\n<ul>\n<li>\n<p>\u7f51\u7edc\u7ec4\u4ef6\u6545\u969c<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u89c2\u5bdf\u5230\u7684\u6240\u6709\u7f51\u7edc\u6545\u969c\u90fd\u53ef\u901a\u8fc7 UFM \u68c0\u6d4b\u5230\uff0c\u5e76\u4f1a\u88ab\u8bb0\u5f55\u5230 UFM \u4e8b\u4ef6\u65e5\u5fd7\u4e2d\uff0c\u56e0\u6b64\u54cd\u5e94\u65b9\u5f0f\u4e5f\u5f88\u7b80\u5355\uff1a\u89e3\u6790 UFM \u65e5\u5fd7\u5e76\u91c7\u53d6\u76f8\u5e94\u63aa\u65bd\u3002<\/p>\n<p>UFM \u4e8b\u4ef6\u7cfb\u7edf\u975e\u5e38\u590d\u6742\uff0c\u5305\u542b\u6570\u5341\u79cd\u4e8b\u4ef6\u7c7b\u578b\u3002\u4f46\u5728\u5b9e\u8df5\u4e2d\uff0c\u6211\u4eec\u53d1\u73b0\u53ea\u6709\u5c11\u6570\u4e8b\u4ef6\u6709\u95ee\u9898\uff0c\u4e3b\u8981\u4e0e\u94fe\u8def\u6545\u969c\u6216\u7b26\u53f7\u9519\u8bef\u6280\u672f\u8f83\u9ad8\u6709\u5173\u3002\u5728\u8bc6\u522b\u51fa\u8fd9\u4e9b\u4e8b\u4ef6\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u7f16\u5199\u811a\u672c\u6765\u89e3\u6790 UFM \u4e8b\u4ef6\u65e5\u5fd7\u3001\u7981\u7528\u4e0e\u6700\u8fd1\u4e8b\u4ef6\u76f8\u5173\u7684\u94fe\u8def\u548c\u7aef\u53e3\u3001\u4e3a\u8fd9\u4e9b\u7f51\u7edc\u7ec4\u4ef6\u7533\u8bf7\u7ef4\u62a4\u5de5\u5355\u3001\u7ef4\u62a4\u5b8c\u6210\u540e\u91cd\u65b0\u542f\u7528\u8fd9\u4e9b\u7ec4\u4ef6\u3002<\/p>\n<ul>\n<li>\n<p>\u672c\u5730\u955c\u50cf\u6587\u4ef6\u7cfb\u7edf<\/p>\n<\/li>\n<\/ul>\n<p>\u5bf9\u4e8e\u8fd9\u4e9b\u5927\u578b\u5206\u5e03\u5f0f\u8bad\u7ec3\uff0c\u4eba\u4eec\u5f88\u65e9\u5c31\u53d1\u73b0\u96c6\u7fa4\u4e0e\u4ee5\u592a\u7f51\u7684\u6570\u636e\u4ea4\u6362\u901f\u5ea6\u662f\u4e00\u5927\u74f6\u9888\u3002\u4e00\u6761\u5171\u4eab\u4ee5\u592a\u7f51\u8fde\u63a5\u7684\u5e26\u5bbd\u5927\u7ea6\u4e3a 10Gbit\/s\uff1b\u5982\u679c\u6709\u6570\u767e\u4e2a\u5de5\u4f5c\u5668\u540c\u65f6\u4e0b\u8f7d\u6570\u636e\u96c6\u548c\u6a21\u578b\u68c0\u67e5\u70b9\uff0c\u90a3\u4e48\u8fd9\u70b9\u5e26\u5bbd\u4f1a\u5f88\u5feb\u9971\u548c\u3002<\/p>\n<p>\u4e3a\u6b64\uff0c\u6211\u4eec\u51b3\u5b9a\u5728\u6211\u4eec\u96c6\u7fa4\u5185\u90e8\u6784\u5efa\u4e00\u4e2a\u672c\u5730\u6587\u4ef6\u7cfb\u7edf\uff0c\u4ee5\u4f5c\u4e3a\u4e91\u5b58\u50a8\u7684\u955c\u50cf\uff0c\u5176\u672c\u8d28\u4e0a\u5c31\u662f\u4e00\u4e2a\u7f13\u5b58\u7a7a\u95f4\uff0c\u53ef\u4ee5\u51cf\u5c11\u4ece S3 \u8bfb\u53d6\u7684\u6587\u4ef6\u91cf\u3002\u4e3a\u4e86\u89e3\u51b3\u96c6\u7fa4\u6d41\u5931\u95ee\u9898\uff08\u5373\u56e0\u4e3a\u7ef4\u4fee\u539f\u56e0\u800c\u7981\u7528\u6216\u66f4\u6362\u673a\u5668\u7684\u60c5\u51b5\uff09\uff0c\u6211\u4eec\u4e3a\u6bcf\u4efd\u6587\u4ef6\u90fd\u51c6\u5907\u4e86\u4e09\u4e2a\u526f\u672c\uff0c\u5e76\u4f7f\u7528\u4e86\u4e00\u81f4\u6027\u54c8\u5e0c\u4ee5\u5747\u5300\u5206\u914d\u8d1f\u8f7d\uff0c\u4ece\u800c\u5728\u96c6\u7fa4\u6d41\u5931\u671f\u95f4\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u6587\u4ef6\u79fb\u52a8\u3002\u7531\u4e8e\u96c6\u7fa4\u7684\u78c1\u76d8\u7a7a\u95f4\u6709\u9650\uff0c\u6240\u4ee5\u6211\u4eec\u5fc5\u987b\u5f00\u53d1\u591a\u79cd\u5de5\u5177\u6765\u8ddf\u8e2a\u6587\u4ef6\u7684\u751f\u547d\u5468\u671f\u548c\u6e05\u9664\u4e0d\u518d\u6709\u7528\u7684\u6587\u4ef6\u3002<\/p>\n<ul>\n<li>\n<p>\u672c\u5730\u5206\u5e03\u5f0f Docker \u6ce8\u518c\u8868<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u4f7f\u7528\u4e86 Kraken\uff0c\u8fd9\u662f\u4e00\u4e2a\u53ef\u70b9\u5bf9\u70b9\u4f20\u8f93 Docker \u955c\u50cf\u7684\u51fa\u8272\u5f00\u6e90\u8f6f\u4ef6\u3002\u8fd9\u4e2a\u8f6f\u4ef6\u51e0\u4e4e\u6ca1\u51fa\u73b0\u8fc7\u4efb\u4f55\u95ee\u9898\uff0c\u6211\u4eec\u8fd8\u662f\u633a\u60ca\u8bb6\u7684\uff0c\u6bd5\u7adf\u6211\u4eec\u7684\u4efb\u52a1\u548c\u5b9e\u73b0\u90fd\u5f88\u590d\u6742\u3002\u5de5\u5177\u5730\u5740\uff1ahttps:\/\/github.com\/uber\/kraken<\/p>\n<ul>\n<li>\n<p>\u5404\u79cd\u6027\u80fd\u76d1\u63a7\u5de5\u5177<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u8bbe\u7f6e\u4e86\u9ed8\u8ba4\u7684 Torch \u5206\u6790\u5668\u4ee5\u53ca\u82f1\u4f1f\u8fbe\u7684 Nsight Systems\u3002\u540e\u8005\u53ef\u5e2e\u52a9\u6211\u4eec\u4e86\u89e3\u524d\u5411 \/ \u53cd\u5411\u901a\u8fc7\u4ee5\u53ca NCCL \u901a\u4fe1\u6240\u9700\u7684\u786e\u5207\u65f6\u95f4\uff0c\u5e76\u8fdb\u4e00\u6b65\u5e2e\u52a9\u6211\u4eec\u786e\u5b9a\u7ed9\u5b9a\u7684\u6a21\u578b\u5927\u5c0f\u548c\u5de5\u4f5c\u5668\u6570\u91cf\u662f\u5426\u4f1a\u6210\u4e3a\u74f6\u9888\u3002\u7136\u800c\uff0cNsight Systems \u6709\u70b9\u96be\u7528\uff0c\u56e0\u4e3a\u5176\u9700\u8981\u5728\u7279\u6743\u6a21\u5f0f\u4e0b\u8fd0\u884c Docker\uff0c\u8fd9\u9700\u8981\u7981\u7528\u4e0e\u6027\u80fd\u76d1\u63a7\u4e8b\u4ef6\u76f8\u5173\u7684\u5b89\u5168\u68c0\u67e5\uff0c\u5e76\u4e14\u4fdd\u5b58\u5176\u914d\u7f6e\u65f6\u5f80\u5f80\u9700\u8981\u505c\u6b62\u6574\u4e2a\u8bad\u7ec3\u8fdb\u7a0b\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u4e5f\u7f16\u5199\u4e86\u5de5\u5177\u6765\u68c0\u6d4b\u8bad\u7ec3\u901f\u5ea6\u6162\u7684\u6570\u636e<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u5e76\u7406\u89e3\u5176\u53ef\u80fd\u539f\u56e0\u3002\u6211\u4eec\u53d1\u73b0\u8fd9\u5f88\u6709\u7528\u3002\u5176\u4e2d\u6700\u6709\u7528\u7684\u5de5\u5177\u7684\u4f5c\u7528\u662f\u76d1\u63a7\u6bcf\u4e00<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u6240\u9700\u7684\u65f6\u95f4\u5e76\u5728\u67d0<mark data-type=\"concepts\" data-id=\"e3995dcb-2c96-4671-b7ed-d5f2abae3c08\">\u6279\u6b21<\/mark>\u8fc7\u4e8e\u6162\u65f6\u4e22\u5f03\u8be5\u5de5\u4f5c\u5668\u7684\u6808\u8ddf\u8e2a \u2014\u2014 \u8fd9\u8ba9\u6211\u4eec\u53ef\u4ee5\u66f4\u8f7b\u677e\u5730\u627e\u5230\u786c\u4ef6\u6216\u8f6f\u4ef6\u6709\u4e9b\u5c0f\u95ee\u9898\u7684\u4e3b\u673a\u3002<\/p>\n<ul>\n<li>\n<p>\u5c06\u673a\u5668\u5206\u6210\u4e0d\u540c\u7684\u7ec4\u522b\u4ee5\u5b9a\u4f4d\u6545\u969c\u4e3b\u673a<\/p>\n<\/li>\n<\/ul>\n<p>\u5728\u4f7f\u7528\u8be5\u96c6\u7fa4\u7684\u524d\u51e0\u4e2a\u6708\uff08\u90a3\u65f6\u5019\u5065\u5eb7\u68c0\u67e5\u8fd8\u4e0d\u5982\u73b0\u5728\u8fd9\u822c\u900f\u5f7b\uff09\uff0c\u6211\u4eec\u7ecf\u5e38\u9047\u5230\u8fd9\u79cd\u60c5\u51b5\uff1a\u5728\u4e00\u7ec4\u673a\u5668\u4e0a\u8bad\u7ec3\u65f6\u51fa\u73b0\u6545\u969c\uff0c\u4f46\u5e76\u4e0d\u6e05\u695a\u7a76\u7adf\u662f\u54ea\u53f0\u673a\u5668\u6709\u95ee\u9898\u3002\u4e3a\u4e86\u627e\u5230\u6545\u969c\u4e3b\u673a\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e9b\u5de5\u5177\uff0c\u53ef\u8f7b\u677e\u5730\u628a\u4e00\u7ec4\u673a\u5668\u5206\u6210\u4e0d\u540c\u7684\u5c0f\u7ec4\uff0c\u7136\u540e\u5728\u6bcf\u4e2a\u673a\u5668\u5c0f\u7ec4\u4e0a\u8fd0\u884c\u66f4\u5c0f\u7684\u8bad\u7ec3\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5982\u679c\u4e00\u4e2a\u5728 48 \u53f0\u673a\u5668\u4e0a\u8fd0\u884c\u7684\u8bad\u7ec3\u5931\u8d25\u4e86\uff0c\u90a3\u4e48\u5c31\u5728 6 \u7ec4\u5404 8 \u53f0\u673a\u5668\u4e0a\u8fdb\u884c\u66f4\u5c0f\u89c4\u6a21\u7684\u8bad\u7ec3\uff0c\u7136\u540e\u5728 8 \u7ec4\u5404 6 \u53f0\u673a\u5668\u4e0a\u8fd0\u884c\u66f4\u5c0f\u89c4\u6a21\u7684\u8bad\u7ec3\u3002\u901a\u5e38\u60c5\u51b5\u4e0b\uff0c\u8fd9\u4e24\u4e2a\u9636\u6bb5\u4e2d\u53ea\u6709\u4e00\u6b21\u8fd0\u884c\u4f1a\u5931\u8d25\uff0c\u8fd9\u8ba9\u6211\u4eec\u6709\u4fe1\u5fc3\u5f97\u51fa\u7ed3\u8bba\uff1a\u5728\u4e24\u4e2a\u9636\u6bb5\u4e2d\u90fd\u51fa\u95ee\u9898\u7684\u673a\u5668\u662f\u6709\u95ee\u9898\u7684\u3002<\/p>\n<p><strong>\u53cd\u601d\u548c\u5b66\u4e60\u5230\u7684\u7ecf\u9a8c\u6559\u8bad<\/strong><\/p>\n<p>\u5728\u8bbe\u7f6e\u548c\u7ef4\u62a4\u57fa\u7840\u8bbe\u65bd\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u83b7\u5f97\u4e86\u4e00\u4e9b\u6709\u7528\u7684\u7ecf\u9a8c\u6559\u8bad\uff1a<\/p>\n<ul>\n<li>\n<p>\u4e00\u79cd\u6709\u7528\u7684\u505a\u6cd5\u662f\u4ea4\u6362\u673a\u5668\u7684\u4f4d\u7f6e\u3002\u5728\u8fd0\u884c\u65f6\uff0c\u4f7f\u7528\u591a\u4e8e\u6240\u9700\u673a\u5668\u6570\u91cf 10-20% \u7684\u673a\u5668\u4f1a\u5f88\u6709\u5e2e\u52a9\uff0c\u8fd9\u6837\u5c31\u80fd\u5728\u673a\u5668\u6545\u969c\u65f6\u8f7b\u677e\u91cd\u542f\u8bad\u7ec3\u4e86\u3002\u8bbe\u7f6e\u96c6\u7fa4\u7f51\u7edc\u8fde\u63a5\u65f6\u8ba9\u6bcf\u53f0\u673a\u5668\u90fd\u4e0e\u5176\u4ed6\u6bcf\u53f0\u673a\u5668\u7d27\u5bc6\u76f8\u8fde\uff0c\u8fd9\u6837\u4e00\u6765\u6211\u4eec\u5c31\u80fd\u4f7f\u7528\u8fd9\u4e9b\u673a\u5668\u4e2d\u4efb\u610f\u53ef\u5de5\u4f5c\u7684\u5b50\u96c6\u3002<\/p>\n<\/li>\n<li>\n<p>\u5bf9\u9047\u5230\u7684\u6bcf\u4e00\u4e2a\u786c\u4ef6\u6216\u8f6f\u4ef6\u6545\u969c\uff0c\u7f16\u5199\u6d4b\u8bd5\u548c\u81ea\u52a8\u5316\u89e3\u51b3\u65b9\u6848\u662f\u503c\u5f97\u7684\uff0c\u56e0\u4e3a\u8bad\u7ec3\u4e2d\u9047\u5230\u7684\u6bcf\u4e00\u4e2a\u95ee\u9898\u90fd\u4f1a\u518d\u6b21\u51fa\u73b0\u3002\u7c7b\u4f3c\u5730\uff0c\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u542b\u6df7\u4e0d\u6e05\u7684\u62a5\u9519\u6d88\u606f\uff0c\u90fd\u503c\u5f97\u7f16\u5199\u66f4\u80fd\u89e3\u91ca\u8be5\u9519\u8bef\u7684\u5de5\u5177\u3002<\/p>\n<\/li>\n<li>\n<p>\u53ef\u91cd\u590d\u6027\u662f\u4f18\u79c0\u79d1\u7814\u7684\u5173\u952e\u3002\u6211\u4eec\u7acb\u9a6c\u5c31\u91c7\u7528\u7684\u4e00\u5927\u539f\u5219\u662f\uff1a\u300c\u4e00\u6b21\u53ea\u4fee\u6539\u4e00\u4e2a\u5730\u65b9\u300d\uff0c\u5373\u4fbf\u6700\u7b80\u5355\u7684\u5730\u65b9\u4e5f\u662f\u5982\u6b64\u3002<\/p>\n<\/li>\n<li>\n<p>\u4fe1\u4efb\uff0c\u4f46\u4e5f\u8981\u9a8c\u8bc1\u3002\u6bcf\u5f53\u6211\u4eec\u5f15\u5165\u5916\u90e8\u5de5\u5177\u6216\u52a0\u5165\u65b0\u4eba\u5458\uff08\u65e0\u8bba\u662f\u6765\u81ea\u516c\u53f8\u5185\u8fd8\u662f\u516c\u53f8\u5916\uff09\u65f6\uff0c\u6211\u4eec\u90fd\u4f1a\u4ed4\u7ec6\u68c0\u67e5\u4ed6\u4eec\u58f0\u79f0\u7684\u4e1c\u897f\uff0c\u5c24\u5176\u662f\u5f53\u540e\u7eed\u6b65\u9aa4\u4f9d\u8d56\u4e8e\u8fd9\u4e9b\u58f0\u79f0\u7684\u4e1c\u897f\u65f6\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u603b\u7ed3<\/strong><\/p>\n<p>\u8bad\u7ec3\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u4e00\u5f00\u59cb\u5c31\u9700\u8981\u590d\u6742\u7684\u57fa\u7840\u8bbe\u65bd\u3002\u6211\u4eec\u4e4b\u6240\u4ee5\u9009\u62e9\u6df1\u5165\u53c2\u4e0e\u57fa\u7840\u8bbe\u65bd\u7684\u8bbe\u7f6e\u7ec6\u8282\uff0c\u662f\u56e0\u4e3a\u6211\u4eec\u76f8\u4fe1\u5b8c\u5168\u7406\u89e3\u6211\u4eec\u64cd\u4f5c\u7684\u7cfb\u7edf\u662f\u975e\u5e38\u91cd\u8981\u7684\uff0c\u4e5f\u56e0\u4e3a\u6211\u4eec\u8ba4\u4e3a\u8fd9\u6837\u505a\u7684\u6548\u7387\u66f4\u9ad8\u3002<\/p>\n<p>\u73b0\u5728\uff0c\u7ecf\u5386\u8fc7\u6574\u4e2a\u6d41\u7a0b\u4e4b\u540e\uff0c\u6211\u4eec\u5f88\u9ad8\u5174\u6211\u4eec\u91c7\u7528\u4e86\u8fd9\u6837\u7684\u65b9\u6cd5 \u2014\u2014 \u4e8b\u5b9e\u8bc1\u660e\uff0c\u80fd\u5b8c\u5168\u63a7\u5236\u6211\u4eec\u7684\u57fa\u7840\u8bbe\u65bd\u4ee5\u53ca\u6709\u80fd\u529b\u8f7b\u677e\u5730\u5728\u6bcf\u4e2a\u62bd\u8c61\u5c42\u7ea7\u4e0a\u8fdb\u884c\u8c03\u8bd5\u5177\u6709\u81f3\u5173\u91cd\u8981\u7684\u4ef7\u503c\u3002\u867d\u7136\u8fd9\u4e2a\u8fc7\u7a0b\u9700\u8981\u5927\u91cf\u7684\u76d1\u7763\u548c\u8fed\u4ee3\uff0c\u4f46\u5b83\u8ba9\u6211\u4eec\u53ef\u4ee5\u6df1\u5165\u5730\u7406\u89e3\u5176\u5e95\u5c42\u5de5\u4f5c\u6d41\u7a0b\u3001\u6784\u5efa\u4e00\u7cfb\u5217\u7528\u4e8e\u786e\u4fdd\u4e3b\u673a\u5065\u5eb7\u7684\u5de5\u5177\u3001\u5b66\u4e60\u5982\u4f55\u8ba9\u7cfb\u7edf\u81ea\u52a8\u8fd0\u884c\u4ee5\u786e\u4fdd\u8bad\u7ec3\u6301\u7eed\u5e73\u6ed1\uff0c\u6700\u7ec8\u6784\u5efa\u8d77\u4e00\u5957\u8ba9\u6211\u4eec\u53ef\u4ee5\u5feb\u901f\u8fed\u4ee3\u8bad\u7ec3\u524d\u6cbf<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n<p>\u8fd9\u4e2a\u57fa\u7840\u8bbe\u65bd\u6784\u5efa\u6d41\u7a0b\u4f53\u73b0\u4e86\u6211\u4eec\u7814\u7a76\u548c\u6784\u5efa AI \u667a\u80fd\u4f53\u7684\u57fa\u7840\u65b9\u6cd5\u8bba\uff1a\u63a2\u7a76\u7ec6\u679d\u672b\u8282\uff0c\u4e0d\u65ad\u6539\u8fdb\u73b0\u6709\u6d41\u7a0b\uff0c\u5e76\u6784\u5efa\u6709\u7528\u7684\u5de5\u5177\u548c\u7cfb\u7edf\u4f7f\u6211\u4eec\u8fd9\u4e2a\u79ef\u6781\u594b\u8fdb\u7684\u56e2\u961f\u80fd\u591f\u5e94\u5bf9\u66f4\u5927\u7684\u6311\u6218\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-24-4\" target=\"_blank\">\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68],"class_list":["post-34716","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=34716\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c\",\"datePublished\":\"2024-12-11T06:57:39+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716\"},\"wordCount\":464,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/frc-ff6478d07102151bce7de146029da7a7.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=34716#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=34716\",\"name\":\"\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/frc-ff6478d07102151bce7de146029da7a7.png\",\"datePublished\":\"2024-12-11T06:57:39+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=34716\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/frc-ff6478d07102151bce7de146029da7a7.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/07\\\/frc-ff6478d07102151bce7de146029da7a7.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34716#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\",\"url\":\"https:\\\/\\\/17aitech.com\\\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/17aitech.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"url\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"contentUrl\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=34716","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=34716#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=34716"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c","datePublished":"2024-12-11T06:57:39+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=34716"},"wordCount":464,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=34716#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ff6478d07102151bce7de146029da7a7.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=34716#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34716","url":"https:\/\/17aitech.com\/?p=34716","name":"\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34716#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34716#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ff6478d07102151bce7de146029da7a7.png","datePublished":"2024-12-11T06:57:39+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34716#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34716"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34716#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ff6478d07102151bce7de146029da7a7.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ff6478d07102151bce7de146029da7a7.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34716#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u4ece\u88f8\u673a\u5230700\u4ebf\u53c2\u6570\u5927\u6a21\u578b\uff0c\u8fd9\u91cc\u6709\u4efd\u6559\u7a0b\uff0c\u8fd8\u6709\u73b0\u6210\u53ef\u7528\u7684\u811a\u672c"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34716","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=34716"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34716\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34716"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34716"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34716"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}