{"id":42205,"date":"2025-12-11T14:31:00","date_gmt":"2025-12-11T06:31:00","guid":{"rendered":"https:\/\/www.yunshan.net\/blog\/?p=42205"},"modified":"2025-12-26T15:43:00","modified_gmt":"2025-12-26T07:43:00","slug":"%e5%9f%ba%e4%ba%8eebpf%e6%8a%80%e6%9c%af%e7%9a%84llm%e6%8e%a8%e7%90%86%e6%9c%8d%e5%8a%a1%e5%85%a8%e6%a0%88%e5%8f%af%e8%a7%82%e6%b5%8b%e6%80%a7%ef%bc%9a%e6%9e%b6%e6%9e%84%e3%80%81%e6%8c%91%e6%88%98","status":"publish","type":"post","link":"https:\/\/www.yunshan.net\/blog\/archives\/42205.html","title":{"rendered":"\u57fa\u4e8eeBPF\u6280\u672f\u7684LLM\u63a8\u7406\u670d\u52a1\u5168\u6808\u53ef\u89c2\u6d4b\u6027\uff1a\u67b6\u6784\u3001\u6311\u6218\u4e0e\u5b9e\u8df5"},"content":{"rendered":"\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u6587\u6863\u6458\u8981\uff1a \u672c\u6587\u7cfb\u7edf\u6027\u9610\u8ff0\u4e86\u5728\u4e91\u539f\u751f\u73af\u5883\u4e0b\uff0c\u5229\u7528\u6269\u5c55\u4f2f\u514b\u5229\u5305\u8fc7\u6ee4\u5668\uff08eBPF\uff09\u6280\u672f\u5b9e\u73b0\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u63a8\u7406\u670d\u52a1\u5168\u6808\u53ef\u89c2\u6d4b\u6027\u7684\u65b9\u6cd5\u4e0e\u67b6\u6784\u3002\u5185\u5bb9\u6db5\u76d6\u81ea\u5efa\u5343\u4ebf\u53c2\u6570LLM\u63a8\u7406\u670d\u52a1\u5728\u5f02\u6784\u786c\u4ef6\uff08\u5982<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=267447924&amp;content_type=Article&amp;match_order=1&amp;q=%E6%98%87%E8%85%BE910B&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">\u6607\u817e910B<\/a>\uff09\u4e0e\u5206\u5e03\u5f0f\u6846\u67b6\uff08\u5982vLLM, Ray\uff09\u4e0b\u9762\u4e34\u7684\u6311\u6218\uff0c\u5206\u6790\u4e86\u4f20\u7edf\u53ef\u89c2\u6d4b\u6027\u4e09\u5927\u652f\u67f1\uff08\u6307\u6807\u3001\u8ffd\u8e2a\u3001\u65e5\u5fd7\uff09\u53ca\u65b0\u5174\u6027\u80fd\u5256\u6790\u652f\u67f1\u7684\u4e0d\u8db3\u3002\u6587\u7ae0\u91cd\u70b9\u4ecb\u7ecd\u4e86DeepFlow\u5e73\u53f0\u5982\u4f55\u901a\u8fc7eBPF\u5b9e\u73b0\u96f6\u4ee3\u7801\u4fb5\u5165\u7684\u5168\u6808\u6307\u6807\u91c7\u96c6\u3001\u5168\u94fe\u8def\u8ffd\u8e2a\u4e0e\u6df7\u5408\u6808\uff08CPU\/GPU\uff09\u6027\u80fd\u5256\u6790\uff0c\u5e76\u8f85\u4ee5\u667a\u80fd\u4f53\u5e94\u7528\u4e0e\u4e2d\u56fd\u79fb\u52a8\u7684\u843d\u5730\u6848\u4f8b\uff0c\u9a8c\u8bc1\u4e86\u8be5\u65b9\u6848\u5728\u4f18\u5316GPU\u5229\u7528\u7387\u3001\u5b9a\u4f4d\u63a8\u7406\u5ef6\u8fdf\u53ca\u9002\u914d\u6d41\u5f0f\u534f\u8bae\u7b49\u65b9\u9762\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<p>\u5173\u952e\u8bcd\uff1a eBPF\uff1bLLM\u63a8\u7406\uff1b\u53ef\u89c2\u6d4b\u6027\uff1b\u5206\u5e03\u5f0f\u8ffd\u8e2a\uff1bGPU\u6027\u80fd\u5256\u6790\uff1bDeepFlow\uff1bvLLM\uff1b\u4e91\u539f\u751f\uff1b<\/p>\n<\/blockquote>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u81ea\u5efaLLM\u63a8\u7406\u670d\u52a1\u7684\u53ef\u89c2\u6d4b\u6027\u6311\u6218<\/h2>\n\n\n\n<p>\u968f\u7740\u5343\u4ebf\u53c2\u6570\u7ea7\u5f00\u6e90\u6a21\u578b\uff08\u5982<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=267447924&amp;content_type=Article&amp;match_order=1&amp;q=DeepSeek&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek<\/a>\uff09\u7684\u666e\u53ca\uff0c\u4f01\u4e1a\u81ea\u5efa\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u63a8\u7406\u670d\u52a1\u7684\u9700\u6c42\u6fc0\u589e\uff0c\u65e0\u8bba\u662f\u4e92\u8054\u7f51\u516c\u53f8\u3001\u884c\u4e1a\u5ba2\u6237\u8fd8\u662f\u653f\u4f01\u5355\u4f4d\uff0c\u90fd\u5f00\u59cb\u79ef\u6781\u642d\u5efa\u81ea\u5df1\u7684\u6a21\u578b\uff0c\u8fd9\u4e5f\u5e26\u52a8\u4e86\u56fd\u4ea7GPU\u7684\u5feb\u901f\u843d\u5730\u3002\u70ed\u6f6e\u4e4b\u4e0b\uff0c\u95ee\u9898\u4e5f\u968f\u4e4b\u6d6e\u73b0\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic1.zhimg.com\/v2-20c63b186d301b11a004c7146123c3d2_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u5982\u4f55\u4fdd\u969c\u81ea\u5efaLLM\u63a8\u7406\u670d\u52a1\u7684\u7528\u6237\u4f53\u9a8c<\/figcaption><\/figure>\n\n\n\n<p>\u4ee5\u5f80\u642d\u5efa\u4e00\u5957\u63a8\u7406\u5f15\u64ce\u6216\u670d\u52a1\u5c1a\u9700\u4e00\u5b9a\u7684\u6280\u672f\u79ef\u7d2f\uff0c\u800cDeepSeek\u7684\u51fa\u73b0\u52a0\u901f\u4e86\u8fd9\u4e00\u8fc7\u7a0b\uff0c\u540c\u65f6\u4e5f\u5e26\u6765\u4e86\u4e24\u65b9\u9762\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u786c\u4ef6\u5f02\u6784\u6027\u4e0e\u8c03\u4f18\u590d\u6742\u5ea6<\/strong>\uff1a\u4e3a\u5e94\u5bf9\u9ad8\u7aefGPU\u7981\u8fd0\uff0c\u5b9e\u8df5\u4e2d\u5e38\u91c7\u7528\u6df7\u5408\u56fd\u4ea7\u5361\uff08\u5982\u534e\u4e3a\u6607\u817e910B\uff09\uff0c\u5176\u89c4\u683c\u3001\u6027\u80fd\u4e0e\u5f00\u53d1\u751f\u6001\uff08\u5982\u6027\u80fd\u5256\u6790\u5de5\u5177\u94fe\uff09\u5dee\u5f02\u663e\u8457\u3002\u6839\u636e\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\uff08MSRA\uff09\u5bf9\u8d85700\u4e2aGPU\u4f4e\u6548\u6848\u4f8b\u7684\u5206\u6790\uff08\u62a5\u544a\u7d22\u5f15\uff1aMSRA-GPU-Eff-2024\uff09\uff0c\u4f4e\u6548\u4e3b\u8981\u539f\u56e0\u5305\u62ec\uff1a\u5185\u6838\u8ba1\u7b97\u914d\u7f6e\u4e0d\u5f53\uff08&gt;20%\uff09\u3001\u663e\u5b58\u64cd\u4f5c\u5ef6\u8fdf\uff08\u7ea625%\uff09\u53ca\u6587\u4ef6I\/O\u74f6\u9888\uff08&gt;10%\uff09\u3002\u5728\u4e0d\u540c\u786c\u4ef6\u4e0a\u590d\u7528\u914d\u7f6e\u5c06\u5bfc\u81f4\u4e25\u91cd\u7684GPU\u5229\u7528\u7387\u4f4e\u4e0b\u3002<\/li>\n\n\n\n<li><strong>\u5206\u5e03\u5f0f\u63a8\u7406\u4e0e\u667a\u80fd\u4f53\u67b6\u6784\u7684\u590d\u6742\u6027<\/strong>\uff1a\u5343\u4ebf\u53c2\u6570\u6a21\u578b\u9700\u8de8\u591a\u673a\u591a\u5361\u8fdb\u884c\u5206\u5e03\u5f0f\u63a8\u7406\uff08\u4f8b\u5982\u57fa\u4e8eACK\u90e8\u7f72\u7684DeepSeek-vLLM-Ray\u67b6\u6784\uff09\u3002\u540c\u65f6\uff0c\u670d\u52a1\u65e5\u76ca\u63a5\u5165\u590d\u6742\u7684AI\u667a\u80fd\u4f53\u6846\u67b6\uff0c\u540e\u8005\u9700\u534f\u8c03\u6a21\u578b\u3001\u5de5\u5177\u4e0e\u8bb0\u5fc6\u7cfb\u7edf\uff0c\u5f62\u6210\u591a\u5c42\u5206\u5e03\u5f0f\u7cfb\u7edf\uff0c\u4f7f\u5f97\u6545\u969c\u5b9a\u754c\u4e0e\u6027\u80fd\u5206\u6790\u6781\u4e3a\u56f0\u96be\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u5728\u56fd\u5185\u7684\u5b9e\u8df5\u4e2d\uff0c\u8fd9\u4e00\u6311\u6218\u5c24\u4e3a\u7a81\u51fa\u3002<strong>\u5bf9\u4e8e\u5927\u578b\u4f01\u4e1a\u800c\u8a00\uff0c\u91c7\u8d2d\u65f6\u5f80\u5f80\u9762\u4e34\u6df7\u5408\u9009\u7528\u4e0d\u540c\u54c1\u724c\u3001\u4e0d\u540c\u578b\u53f7\u5f02\u6784\u5361\u7684\u590d\u6742\u60c5\u51b5\u3002\u5982\u4f55\u8fdb\u884c\u9009\u578b\u3001\u914d\u6bd4\uff0c\u4ee5\u53ca\u540e\u7eed\u7684\u53c2\u6570\u8c03\u4f18\uff0c\u90fd\u6210\u4e86\u4e9f\u5f85\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/strong><\/p>\n\n\n\n<p><strong>\u90a3\u4e48\u4e3a\u4f55\u53c2\u6570\u8c03\u4f18\u5982\u6b64\u5173\u952e\uff1f<\/strong>\u6839\u636e\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\uff08MSRA\uff09\u4e8e2024\u5e74\u53d1\u5e03\u7684\u300aGPU Efficiency Analysis in LLM Inference\u300b\u62a5\u544a\uff08\u53c2\u8003\u7f16\u53f7 MSRA-2024-GPU-Eff\uff09\u66fe\u5bf9\u4e03\u767e\u591a\u4e2aGPU\u4f4e\u6548\u4f7f\u7528\u6848\u4f8b\u8fdb\u884c\u5206\u6790\uff0c\u53d1\u73b0\u4f4e\u6548\u60c5\u51b5\u5927\u81f4\u53ef\u5f52\u7eb3\u4e3a\u4e09\u7c7b\uff0c\u5404\u5360\u7ea6\u56db\u5206\u4e4b\u4e00\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic3.zhimg.com\/v2-463105c81f6db7dda79b0bd5bb0941ea_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u540c\u6837\u7684\u8bad\u7ec3\/\u63a8\u7406\u4ee3\u7801\u5728\u4e0d\u540c\u786c\u4ef6\u4e0b\u7684\u6027\u80fd\u8868\u73b0<\/figcaption><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7b2c\u4e00\u7c7b\u95ee\u9898\uff1a\u8ba1\u7b97GPU\u5185\u6838<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u5728\u6d4b\u8bd5\u96c6\u7fa4\u4e2d\u4f7f\u7528\u8f83\u5c0f\u7684batch size\u8fdb\u884c\u8bad\u7ec3\u6216\u63a8\u7406\uff0c\u4ee5\u9002\u5e94\u6d4b\u8bd5\u73af\u5883\uff1b\u4f46\u5230\u4e86\u751f\u4ea7\u96c6\u7fa4\uff0c\u82e5\u672a\u6839\u636e\u4e0d\u540c\u7684\u786c\u4ef6\u6761\u4ef6\u8c03\u6574\u6b64\u53c2\u6570\uff0c\u5c31\u53ef\u80fd\u5bfc\u81f4GPU\u5229\u7528\u7387\u4f4e\u4e0b\u2014\u2014\u5728A\u786c\u4ef6\u4e0a\u6709\u6548\u7684\u914d\u7f6e\uff0c\u5728B\u786c\u4ef6\u4e0a\u53ef\u80fd\u9700\u8981\u5b8c\u5168\u4e0d\u540c\u3002\u8fd9\u7c7b\u56e0\u4ee3\u7801\u914d\u7f6e\u5bfc\u81f4\u7684\u95ee\u9898\uff0c\u7ea6\u5360\u4f4e\u6548\u6848\u4f8b\u768420%\u4ee5\u4e0a\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7b2c\u4e8c\u7c7b\u95ee\u9898\uff1a\u663e\u5b58\u64cd\u4f5c<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u4ee3\u7801\u4e2d\u5c06\u6570\u636e\u4ece\u4e3b\u673a\u5185\u5b58\u590d\u5236\u5230HBM\u663e\u5b58\uff0c\u6216\u5728\u4e0d\u540c\u5361\u4e4b\u95f4\u8fdb\u884c\u6570\u636e\u62f7\u8d1d\u65f6\uff0c\u53ef\u80fd\u56e0\u672a\u59a5\u5584\u5904\u7406\u963b\u585e\u64cd\u4f5c\u800c\u9020\u6210\u4e0d\u5fc5\u8981\u7684\u5ef6\u8fdf\u3002\u8fd1\u671fDeepSeek\u5728\u5f00\u6e90\u5468\u62a5\u4e2d\u53d1\u5e03\u7684\u6027\u80fd\u5256\u6790\u6570\u636e\u4e5f\u81f4\u529b\u4e8e\u4f18\u5316\u6b64\u7c7b\u901a\u4fe1\u4e0e\u8ba1\u7b97\u7684\u91cd\u53e0\u6267\u884c\uff0c\u4ee5\u5c3d\u53ef\u80fd\u63d0\u5347GPU\u5229\u7528\u7387\u3002\u8fd9\u540c\u6837\u662f\u5bfc\u81f4GPU\u4f4e\u6548\u8fd0\u8f6c\u7684\u539f\u56e0\u4e4b\u4e00\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7b2c\u4e09\u7c7b\u95ee\u9898\uff1a\u6587\u4ef6\u8bfb\u5199<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u5728\u8bad\u7ec3\u4efb\u52a1\u4e2d\uff0c\u5c06\u68c0\u67e5\u70b9\u5199\u5165\u672c\u5730\u78c1\u76d8\u6216\u5206\u5e03\u5f0f\u6587\u4ef6\u7cfb\u7edf\u65f6\uff0c\u82e5I\/O\u6027\u80fd\u4e0d\u8db3\uff0c\u4f1a\u5bf9\u6574\u4e2a\u8bad\u7ec3\u8fdb\u7a0b\u9020\u6210\u74f6\u9888\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic4.zhimg.com\/v2-afd638b48d448bb62c1267c03339c7dd_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u5343\u4ebf\u53c2\u6570LLM\u63a8\u7406\u670d\u52a1\u3001AI\u667a\u80fd\u4f53\u70b9\u590d\u6742\u6027<\/figcaption><\/figure>\n\n\n\n<p>\u56de\u5230\u5343\u4ebf\u53c2\u6570\u6a21\u578b\u7684\u63a8\u7406\u573a\u666f\u3002\u5982\u56fe\uff08\u5343\u4ebf\u53c2\u6570LLM\u63a8\u7406\u670d\u52a1\u3001AI\u667a\u80fd\u4f53\u70b9\u590d\u6742\u6027\uff09\u6240\u793a\uff0c\u5de6\u4fa7\u793a\u610f\u5728\u5343\u4ebf\u53c2\u6570\u89c4\u6a21\u4e0b\uff0c\u5f80\u5f80\u4e0d\u5f97\u4e0d\u91c7\u7528\u5206\u5e03\u5f0f\u63a8\u7406\u670d\u52a1\u3002\u4ee5\u5b9e\u73b0\u591a\u8282\u70b9\u3001\u591a\u5361\u4e0a\u5343\u4ebf\u53c2\u6570\u6a21\u578b\u7684\u6709\u6548\u63a8\u7406\u3002<\/p>\n\n\n\n<p>\u8fd9\u8fd8\u53ea\u662f\u76f8\u5bf9\u7b80\u5355\u7684\u5206\u5e03\u5f0f\u5f62\u6001\uff0c\u4ec5\u6d89\u53ca\u82e5\u5e72\u8fdb\u7a0b\u3002\u53f3\u4fa7\u5219\u5c55\u793a\u4e86\u53e6\u4e00\u4e2a\u65e5\u76ca\u5e38\u89c1\u7684\u573a\u666f\uff1a<strong>\u63a8\u7406\u670d\u52a1\u9700\u8981\u63a5\u5165\u5f53\u524d\u70ed\u95e8\u7684AI\u667a\u80fd\u4f53\u4f53\u7cfb\u3002<\/strong>Chatbot\u53ea\u662f\u4e00\u79cd\u7b80\u5355\u5f62\u6001\uff0c\u82e5\u8003\u8651\u5b8c\u6574\u7684AI\u667a\u80fd\u4f53\u5e94\u7528\uff0c\u5b83\u8fd8\u9700\u8981\u4e0e\u6a21\u578b\u3001\u5de5\u5177\u4ee5\u53ca\u8bb0\u5fc6\u7cfb\u7edf\u7b49\u8fdb\u884c\u4ea4\u4e92\u3002\u8fd9\u5b9e\u9645\u4e0a\u6784\u6210\u4e86\u4e00\u4e2a\u76f8\u5f53\u590d\u6742\u7684\u5206\u5e03\u5f0f\u7cfb\u7edf\u3002<\/p>\n\n\n\n<p>\u603b\u7684\u6765\u8bf4\uff0c\u4ee5\u4e0a\u6240\u8ff0\u6b63\u662f\u5f53\u524d\u63a8\u7406\u670d\u52a1\u9762\u4e34\u7684\u591a\u91cd\u6311\u6218\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001LLM\u63a8\u7406\u670d\u52a1\u53ef\u89c2\u6d4b\u6027\u7684\u6838\u5fc3\u652f\u67f1\u53ca\u4f20\u7edf\u65b9\u6848\u7684\u5c40\u9650<\/h2>\n\n\n\n<p>\u8c08\u5230\u53ef\u89c2\u6d4b\u6027\uff0c\u5927\u5bb6\u901a\u5e38\u90fd\u4f1a\u63d0\u53ca\u7ecf\u5178\u7684\u4e09\u5927\u652f\u67f1\uff1a\u6307\u6807\u3001\u8ffd\u8e2a\u4e0e\u65e5\u5fd7\u3002\u7406\u8bba\u4e0a\uff0c\u5c06\u8fd9\u4e9b\u6570\u636e\u6536\u96c6\u9f50\u5168\u4fbf\u8db3\u4ee5\u652f\u6491\u89c2\u6d4b\u9700\u6c42\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic4.zhimg.com\/v2-6fa7b07129e48ba9a20a372e613194d9_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u4e91\u539f\u751f\u73af\u5883\u4e0bLLM\u63a8\u7406\u670d\u52a1\u7684\u53ef\u89c2\u6d4b\u6027<\/figcaption><\/figure>\n\n\n\n<p>\u800cOTel\u57282024\u5e74\u5c06\u6027\u80fd\u5256\u6790\u6570\u636e\u63d0\u5347\u4e3a\u81f3\u5173\u91cd\u8981\u7684\u7b2c\u56db\u5927\u652f\u67f1\uff0c\u96c6\u9f50\u8fd9\u4e9b\u80fd\u529b\u4f3c\u4e4e\u5c31\u591f\u4e86\uff1f\u5b9e\u9645\u4e0a\uff0cLLM\u63a8\u7406\u5f15\u64ce\u6216\u63a8\u7406\u670d\u52a1\u66f4\u8d34\u8fd1\u57fa\u7840\u8bbe\u65bd\u5c42\uff0c\u8fd9\u4e0e\u901a\u7528\u7684\u5206\u5e03\u5f0f\u5e94\u7528\u7cfb\u7edf\u6709\u6240\u4e0d\u540c\u3002\u5f53\u524d\u8bb8\u591a\u5b9e\u8df5\u4e5f\u5efa\u8bae\u5728&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=267447924&amp;content_type=Article&amp;match_order=1&amp;q=K8s&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">K8s<\/a>&nbsp;\u5bb9\u5668\u73af\u5883\u4e2d\u642d\u5efa\u63a8\u7406\u670d\u52a1\uff0c\u8fd9\u4f7f\u5f97\u53ef\u89c2\u6d4b\u6027\u9762\u4e34\u4e00\u4e9b\u5dee\u5f02\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6307\u6807<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u4f8b\u5982\u5728\u6307\u6807\u5c42\u9762\uff0cvLLM\u4e0eRay\u7b49\u8fdb\u7a0b\u4e4b\u95f4\u5728\u63a8\u7406\u65f6\uff0cToken\u662f\u4ee5\u6d41\u5f0f\u65b9\u5f0f\u8f93\u51fa\u7684\u3002\u89c2\u6d4b\u201cTTFT\u201d\uff08Time to First Token\uff0c\u9996\u6b21\u8f93\u51faToken\u7684\u65f6\u95f4\uff09\u6216\u201cTPOT\u201d\uff08Time Per Output Token\uff0c\u6bcf\u6b21\u8f93\u51faToken\u7684\u65f6\u95f4\uff09\u65f6\uff0c\u4e0d\u80fd\u53ea\u5173\u6ce8\u5355\u4e00\u8282\u70b9\uff0c\u800c\u9700\u8981\u5ba1\u89c6\u5168\u94fe\u8def\u7684\u6bcf\u4e00\u4e2a\u73af\u8282\uff0c\u6bd4\u5982K8s\u4e2d\u6bcf\u4e00\u5f20\u7f51\u5361\u7684\u65f6\u8fdf\u6d88\u8017\u60c5\u51b5\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8ffd\u8e2a<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u5728\u8ffd\u8e2a\u65b9\u9762\uff0c\u5bf9\u4e8eK8s\u57fa\u7840\u8bbe\u65bd\uff0c\u6211\u4eec\u9664\u4e86\u8981\u770b\u5230\u6bcf\u4e2a\u73af\u8282\u7684\u65f6\u5ef6\uff0c\u8fd8\u9700\u5173\u6ce8\u57fa\u7840\u8bbe\u65bd\u670d\u52a1\uff08\u5982\u6700\u57fa\u7840\u7684DNS\uff0c\u6216\u6570\u636e\u5e93\u3001\u5411\u91cf\u6570\u636e\u5e93\u7b49\uff09\u5728\u6574\u4e2a\u8c03\u7528\u94fe\u4e2d\u7684\u72b6\u6001\u4e0e\u6027\u80fd\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6027\u80fd<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u9664\u4e86CPU\u5256\u6790\uff0c\u66f4\u9700\u8981\u4e0eGPU\u4e0a\u4e0b\u6587\u7ed3\u5408\uff1b\u9664\u4e86\u4e3b\u5185\u5b58\u7684\u5256\u6790\uff0c\u8fd8\u9700\u8981\u5177\u5907\u9ad8\u5e26\u5bbd\u5185\u5b58\uff08\u5373\u663e\u5b58\uff09\u7684\u5256\u6790\u80fd\u529b\u3002\u8fd9\u80fd\u5e2e\u52a9\u6211\u4eec\u5feb\u901f\u8bc6\u522b\u63a8\u7406\u5f15\u64ce\u8fd0\u884c\u65f6\u7684OOM\uff08Out of Memory\uff09\u6839\u6e90\u3002\u4e8b\u5b9e\u4e0a\uff0c\u5728vLLM\u7684GitHub Issue\u4e2d\uff0c\u65f6\u5e38\u51fa\u73b0\u5347\u7ea7\u65b0\u7248\u672c\u6216\u4f7f\u7528\u67d0\u578b\u53f7\u5361\u65f6\u5f15\u53d1\u7684\u663e\u5b58OOM\u95ee\u9898\uff0c\u8fd9\u4e5f\u662f\u4e00\u4e2a\u65b0\u7684\u6311\u6218\u3002<\/p>\n\n\n\n<p><strong>\u4f9d\u9760\u4f20\u7edf\u624b\u6bb5\u53ef\u4ee5\u5b8c\u7f8e\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\u548c\u95ee\u9898\u5417\uff1f\u5b9e\u9645\u4e0a\u4f9d\u9760\u4f20\u7edf\u624b\u6bb5\u65e0\u8bba\u4efb\u4f55\u65b9\u5f0f\uff0c\u90fd\u4f1a\u5b58\u5728\u4e00\u5b9a\u7684\u5c40\u9650\u3002\u4f8b\u5982\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>2.1 \u63a8\u7406\u670d\u52a1\u57fa\u7840\u8bbe\u65bd\u6307\u6807\u6570\u636e<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic2.zhimg.com\/v2-07c0cc97e48c95509d39154ceaa657f7_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u63a8\u7406\u670d\u52a1Infra\u6307\u6807\uff1aPrometheus+NVIDIA DCGM Exporter<\/figcaption><\/figure>\n\n\n\n<p>\u6307\u6807\u901a\u5e38\u5206\u4e3a\u57fa\u7840\u8bbe\u65bd\u6307\u6807\u548c\u4e1a\u52a1\u6307\u6807\u3002\u57fa\u7840\u8bbe\u65bd\u6307\u6807\u4e3b\u8981\u6307GPU\u5361\u7684\u7528\u91cf\u3001\u529f\u8017\u7b49\uff0c\u8fd9\u9700\u8981\u7531\u786c\u4ef6\u4e3b\u52a8\u66b4\u9732\u3002<strong>\u4f46\u8fd9\u91cc\u5b58\u5728\u4e00\u4e2a\u5c40\u9650\uff1a<\/strong>\u5982\u679cGPU\u88ab\u865a\u62df\u5316\uff0c\u6bcf\u4e2a\u865a\u62df\u5361\u7684\u8be6\u7ec6\u6307\u6807\u53d7\u9650\u4e8e\u786c\u4ef6\u672c\u8eab\u7684\u66b4\u9732\u80fd\u529b\uff0c\u5f88\u53ef\u80fd\u65e0\u6cd5\u5b8c\u6574\u83b7\u53d6\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>2.2 \u63a8\u7406\u670d\u52a1\u4e1a\u52a1\u6307\u6807\u6570\u636e<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/picx.zhimg.com\/v2-c2e8e67de0884c9ed6cd810f9b5040f7_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u63a8\u7406\u670d\u52a1\u4e1a\u52a1\u6307\u6807\uff1aPrometheus+L7 AI Gateway Metrics<\/figcaption><\/figure>\n\n\n\n<p>\u4e1a\u52a1\u6307\u6807\u5982\u679c\u4f9d\u8d56\u63a8\u7406\u5f15\u64ce\u81ea\u8eab\u63d0\u4f9b\uff0c\u5176\u5b8c\u5907\u6027\u53ef\u80fd\u4e0d\u8db3\u3002\u5f53\u524d\u5f00\u6e90\u7684\u63a8\u7406\u5f15\u64ce\u4f17\u591a\uff0c\u672a\u5fc5\u80fd\u4e3a\u5404\u79cd\u6a21\u578b\u63d0\u4f9b\u4e30\u5bcc\u7684\u4e1a\u52a1\u6307\u6807\u3002\u4e00\u79cd\u5e38\u89c1\u7684\u53d8\u901a\u65b9\u6848\u662f\u901a\u8fc7\u7f51\u7edc\u7f51\u5173\u8fdb\u884c\u89c2\u6d4b\u3002\u76ee\u524d\u8bb8\u591aLLM\u7f51\u5173\u90fd\u5177\u5907AI Gateway\u80fd\u529b\uff0c\u53ef\u4ee5\u5728\u7f51\u5173\u4fa7\u66b4\u9732\u4e1a\u52a1\u6307\u6807\u3002<strong>\u4f46\u65e0\u8bba\u54ea\u79cd\u65b9\u6cd5\u90fd\u4f1a\u5b58\u5728\u95ee\u9898\uff1a<\/strong>\u82e5\u7531\u4e1a\u52a1\u81ea\u8eab\u66b4\u9732\uff0c\u5219\u9700\u8981\u63d2\u5165Prometheus SDK\u8fdb\u884c\u57cb\u70b9\uff1b\u82e5\u4ec5\u5728\u7f51\u5173\u5904\u66b4\u9732\uff0c\u5219\u89c2\u6d4b\u89c6\u89d2\u4ec5\u9650\u4e8e\u7f51\u5173\u8fd9\u4e00\u4e2a\u70b9\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>2.3 \u8c03\u7528\u8fde\u8ffd\u8e2a<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic4.zhimg.com\/v2-bd824c0d53154b85fd8c8b21d5b8b75b_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u8c03\u7528\u94fe\u8ffd\u8e2a\uff1a\u5206\u5e03\u5f0f\u63a8\u7406\u670d\u52a1 Instrumentation<\/figcaption><\/figure>\n\n\n\n<p>\u7531\u4e8e\u63a8\u7406\u670d\u52a1\u53ca\u5176\u524d\u7f6e\u7684\u667a\u80fd\u4f53\u6216\u63a5\u5165\u670d\u52a1\u6784\u6210\u4e86\u4e00\u4e2a\u590d\u6742\u7684\u5206\u5e03\u5f0f\u5e94\u7528\uff0c\u6784\u5efa\u94fe\u8def\u8ffd\u8e2a\u80fd\u529b\u5c31\u663e\u5f97\u975e\u5e38\u5173\u952e\u3002\u76ee\u524d\u5df2\u6709\u4e0d\u5c11\u5f00\u6e90\u5de5\u5177\uff0c\u4f8b\u5982\u57fa\u4e8e<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=267447924&amp;content_type=Article&amp;match_order=1&amp;q=OpenTelemetry&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">OpenTelemetry<\/a>\u7684OpenTracing\uff0c\u4ee5\u53caOpenLIT\u7b49\uff0c\u5b83\u4eec\u652f\u6301\u7684\u80fd\u529b\u76f8\u8fd1\uff0c\u4e3b\u8981\u901a\u8fc7Python\u6216JavaScript\/TypeScript\u7684\u63d2\u6869\u5e93\u6765\u5b9e\u73b0\u5206\u5e03\u5f0f\u8ffd\u8e2a\u3002<strong>\u4f46\u73b0\u6709\u65b9\u6848\u4e5f\u5b58\u5728\u4e00\u4e9b\u4e0d\u8db3\uff1a<\/strong>\u5176\u4e00\uff0c\u901a\u5e38\u53ea\u80fd\u5173\u6ce8\u5355\u4e2a\u8fdb\u7a0b\u5185\u90e8\u7684\u8017\u65f6\u6d88\u8017\uff0c\u96be\u4ee5\u6269\u5c55\u5230\u6574\u4e2aK8s\u94fe\u8def\u4e0a\uff0c\u5448\u73b0\u4ece\u57fa\u7840\u8bbe\u65bd\u5230\u5e94\u7528\u7684\u5168\u8def\u5f84\u65f6\u5ef6\u3002\u5176\u4e8c\uff0c\u5bf9\u591a\u79cd\u7f16\u7a0b\u8bed\u8a00\u7684\u652f\u6301\u4ecd\u5728\u9010\u6b65\u5b8c\u5584\u4e2d\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>2.4 \u6027\u80fd\u5256\u6790<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic4.zhimg.com\/v2-c789ba61aa589e32a60fc9376b7dce8d_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u6027\u80fd\u5256\u6790\uff1aGPU Profiling\u3001HBM Porfiling<\/figcaption><\/figure>\n\n\n\n<p>\u5982\u56fe\uff08\u6027\u80fd\u5256\u6790\uff1aGPU Profiling\u3001HBM Porfiling\uff09\u6240\u793a\uff0c\u5de6\u4fa7\u662f<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=267447924&amp;content_type=Article&amp;match_order=1&amp;q=NVIDIA+Nsight&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">NVIDIA Nsight<\/a>\u63d0\u4f9b\u7684\u6027\u80fd\u5256\u6790\u80fd\u529b\u3002\u8fd9\u662f\u4e00\u79cd\u8f83\u4e3a\u5e95\u5c42\u7684\u5de5\u5177\uff0c\u4e3b\u8981\u9762\u5411\u7cfb\u7edf\u5de5\u7a0b\u5e08\u6216GPU\u5de5\u7a0b\u5e08\u8fdb\u884c\u6027\u80fd\u8c03\u4f18\uff0c\u53ef\u4ee5\u6df1\u5165\u89c2\u5bdfGPU\u5185\u90e8\u5904\u7406\u5355\u5143\u6216\u663e\u5b58\u7684\u5229\u7528\u7387\u7b49\u60c5\u51b5\u3002\u4f46\u5bf9\u4e8e\u4e0a\u5c42\u5e94\u7528\u7684\u5f00\u53d1\u8005\u800c\u8a00\uff0c\u8fd9\u7c7b\u5de5\u5177\u7f3a\u5c11CPU\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u5b58\u5728\u4e00\u5b9a\u8ba4\u77e5\u9694\u9602\u3002<\/p>\n\n\n\n<p>\u4ee5\u5f80\uff0c\u63a8\u7406\u6216\u8bad\u7ec3\u4efb\u52a1\u53ef\u80fd\u53ea\u7531\u5c11\u6570\u4e13\u4e1a\u5de5\u7a0b\u5e08\u8d1f\u8d23\uff0c\u4ed6\u4eec\u80fd\u591f\u89e3\u8bfb\u5de6\u4fa7\u7684\u5256\u6790\u6570\u636e\u3002\u7136\u800c\u5f53\u4e0b\uff0c\u8d8a\u6765\u8d8a\u591a\u7684\u5de5\u7a0b\u5e08\u9700\u8981\u53c2\u4e0e\u63a8\u7406\u670d\u52a1\u7684\u5f00\u53d1\u548c\u8fd0\u7ef4\uff0c\u4ed6\u4eec\u66f4\u719f\u6089\u4e14\u9700\u8981\u7684\u662f\u8d34\u8fd1CPU\u4e0a\u4e0b\u6587\u3001\u80fd\u591f\u5c55\u793aPython\u51fd\u6570\u8c03\u7528\u6808\u7684\u5256\u6790\u80fd\u529b\u2014\u2014\u8fd9\u6b63\u662f\u53f3\u4fa7<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=267447924&amp;content_type=Article&amp;match_order=1&amp;q=PyTorch+Profiler&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">PyTorch Profiler<\/a>\u6240\u63d0\u4f9b\u7684\u3002\u4e14\u53f3\u4e0b\u89d2\u5c55\u793a\u7684\u51fd\u6570\u8c03\u7528\u6808\u4fe1\u606f\uff0c\u5bf9\u4e8e\u5728GPU\u4e4b\u4e0a\u8fdb\u884c\u5e94\u7528\u5f00\u53d1\u7684\u5de5\u7a0b\u5e08\u6765\u8bf4\u66f4\u52a0\u76f4\u89c2\u53cb\u597d\u3002<\/p>\n\n\n\n<p><strong>\u7136\u800c\u5728Python\u4e2d\u83b7\u53d6\u6b64\u7c7b\u8c03\u7528\u6808\u4fe1\u606f\uff0c\u901a\u5e38\u9700\u8981\u8fdb\u884c\u624b\u5de5\u6253\u9020\u7684\u7cbe\u7ec6\u63d2\u6869\u3002<\/strong>\u4f8b\u5982\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u53ef\u80fd\u9700\u8981\u8bbe\u7f6e\u7b49\u5f85\u8f6e\u6b21\u3001\u9884\u70ed\u8f6e\u6b21\u3001\u6fc0\u6d3b\u8f6e\u6b21\u7b49\u3002\u7531\u4e8e\u6027\u80fd\u5f00\u9500\u8f83\u5927\uff0c\u8fd9\u79cd\u63d2\u6869\u9700\u8981\u7cbe\u5de7\u7684\u8bbe\u8ba1\uff1a\u4e00\u822c\u5728\u6d4b\u8bd5\u9636\u6bb5\u5148\u8fd0\u884c\u4e00\u6b21\u4ee5\u83b7\u53d6\u5256\u6790\u6570\u636e\uff0c\u636e\u6b64\u8fdb\u884c\u4f18\u5316\uff0c\u800c\u540e\u5728\u6b63\u5f0f\u8bad\u7ec3\u6216\u63a8\u7406\u524d\u79fb\u9664\u8fd9\u90e8\u5206\u4ee3\u7801\u3002<\/p>\n\n\n\n<p>\u7efc\u4e0a\u6240\u8ff0\uff0c<strong>\u4f20\u7edf\u624b\u6bb5<\/strong>\u80fd\u8986\u76d6\u6307\u6807\u3001\u8ffd\u8e2a\u3001\u5256\u6790\u548c\u65e5\u5fd7\u8fd9\u56db\u5927\u652f\u67f1\uff0c<strong>\u4f46\u5728\u4e24\u79cd\u573a\u666f\u4e0b\u4ecd\u663e\u4e0d\u8db3\uff1a<\/strong>\u7b2c\u4e00\uff0c\u968f\u7740\u66f4\u591a\u5e94\u7528\u5de5\u7a0b\u5e08\u9700\u8981\u9762\u5bf9\u63a8\u7406\u670d\u52a1\uff0c\u5de5\u5177\u9700\u8981\u53d8\u5f97\u66f4\u7b80\u5355\u6613\u7528\uff1b\u7b2c\u4e8c\uff0c\u5728\u4e91\u539f\u751f\u73af\u5883\u4e2d\u53ea\u89c2\u6d4b\u5355\u70b9\uff0c\u5728\u6392\u67e5\u95ee\u9898\u65f6\u53ef\u80fd\u5ffd\u7565\u67d0\u4e9b\u7531\u57fa\u7840\u8bbe\u65bd\uff0c\u6216\u8005\u7279\u5b9a\u578b\u53f7GPU\u5361\u5bfc\u81f4\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001DeepFlow \u57fa\u4e8eeBPF\u7684\u5168\u6808\u53ef\u89c2\u6d4b\u6027\u89e3\u51b3\u65b9\u6848<\/h2>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic2.zhimg.com\/v2-f8407494e510074d4825d42108010645_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u4e3a\u4ec0\u4e48\u4f7f\u7528eBPF\uff1a\u96f6\u4fb5\u6270\u3001\u5168\u6808<\/figcaption><\/figure>\n\n\n\n<p>\u5982\u56fe\uff08\u4e3a\u4ec0\u4e48\u4f7f\u7528eBPF\uff1a\u96f6\u4fb5\u6270\u3001\u5168\u6808\uff09\u6240\u793a\uff0c\u5de6\u4fa7\u793a\u610f\u56fe\u89c1\u4e8e<a href=\"https:\/\/link.zhihu.com\/?target=http%3A\/\/eBPF.io\" target=\"_blank\" rel=\"noreferrer noopener\">http:\/\/eBPF.io<\/a>\u5b98\u7f51\uff0c\u5176\u6838\u5fc3\u4f18\u52bf\u53ef\u6982\u62ec\u4e3a\u201c\u96f6\u4ee3\u7801\u4fee\u6539\u201d\u4e0e\u201c\u5168\u6808\u89c2\u6d4b\u201d\u3002\u5373\u65e0\u9700\u6539\u52a8\u5e94\u7528\u4ee3\u7801\u6216\u91cd\u542f\u8fdb\u7a0b\uff0c\u5373\u53ef\u83b7\u53d6\u53ef\u89c2\u6d4b\u6027\u6570\u636e\uff1b\u5e76\u4e14\u4e0d\u4ec5\u80fd\u83b7\u5f97\u8fdb\u7a0b\u5185\u90e8\u7684\u6570\u636e\uff0c\u8fd8\u80fd\u8986\u76d6Linux\u5185\u6838\u4e43\u81f3\u6574\u4e2a\u8f6f\u4ef6\u6808\u7684\u89c2\u6d4b\u4fe1\u606f\u3002<\/p>\n\n\n\n<p>\u53f3\u4fa7\u662feBPF\u5728\u7528\u6237\u6001\u540c\u6837\u5177\u5907\u5f3a\u5927\u7684uprobe\u80fd\u529b\u3002\u5c3d\u7ba1\u5e38\u6709\u4eba\u8bf4uprobe\u7684\u5f00\u9500\u4f1a\u6bd4\u5185\u6838\u6001\u7684kprobe\u66f4\u5927\uff0c\u4f46\u8fd9\u91cc\u6709\u4e00\u4e2a\u5173\u952e\u70b9\uff1a\u5728\u8ba1\u7b97\u5bc6\u96c6\u578b\u573a\u666f\u4e2d\uff0c\u8bb8\u591a\u5173\u952e\u64cd\u4f5c\uff08\u5982GPU\u5185\u6838\u51fd\u6570\u8c03\u7528\uff0c\u6216\u57fa\u4e8eInfiniBand\u7b49\u9ad8\u901f\u7f51\u7edc\u7684\u901a\u4fe1\u8c03\u7528\uff09\u672c\u8eab\u5f00\u9500\u5c31\u5df2\u975e\u5e38\u5927\u3002\u56e0\u5176\u81ea\u8eab\u5f00\u9500\u5de8\u5927\uff0c\u5728\u5176\u4e0a\u9644\u52a0uprobe\u6240\u5e26\u6765\u7684\u989d\u5916\u8d44\u6e90\u6d88\u8017\u6216\u65f6\u95f4\u4ee3\u4ef7\u5f80\u5f80\u5fae\u4e0d\u8db3\u9053\u3002\u6839\u636e DeepFlow \u7684\u5b9e\u9645\u6d4b\u7b97\uff0c\u989d\u5916\u635f\u8017\u4ec5\u5728\u4e2a\u4f4d\u6570\u767e\u5206\u6bd4\uff08\u4f8b\u59821%\u52305%\uff09\u4e4b\u95f4\u3002\u56e0\u6b64\uff0c\u5728\u53f3\u4fa7\u6240\u793a\u7684\u8fd9\u7c7b\u8ba1\u7b97\u573a\u666f\u4e2d\uff0cuprobe\u662f\u5b8c\u5168\u53ef\u7528\u7684\uff0c\u5176\u5f00\u9500\u6709\u65f6\u751a\u81f3\u4f4e\u4e8e\u8bb8\u591a\u4f20\u7edf\u63d2\u6869\u65b9\u5f0f\u5f15\u5165\u7684\u6027\u80fd\u635f\u8017\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>3.1 DeepFlow \u5168\u6808\u6307\u6807\u91c7\u96c6\u89e3\u51b3\u65b9\u6848<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/picx.zhimg.com\/v2-555a71082b824871a4a7d9f72c1c8675_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">Metrics\uff1a\u4e91\u539f\u751f\u73af\u5883\u4e0b\u9700\u8981\u5168\u6808\u6027\u80fd\u6307\u6807<\/figcaption><\/figure>\n\n\n\n<p>\u4ee5\u4e00\u4e2a\u5178\u578b\u573a\u666f\u4e3a\u4f8b\uff1a\u5047\u8bbe\u4e00\u4e2aNginx Gateway\u5c06\u8d1f\u8f7d\u5206\u53d1\u5230\u540e\u7aef\u7684Serving Pod\u3002\u5982\u679c\u51fa\u73b0TTFT\u6216TPOT\u6307\u6807\u7684\u5f02\u5e38\u6ce2\u52a8\uff0c\u800c\u6211\u4eec\u65e0\u6cd5\u786e\u5b9a\u95ee\u9898\u662f\u5426\u51fa\u5728\u6a21\u578b\u670d\u52a1\u4fa7\uff0c\u90a3\u4e48\u4e2d\u95f4K8s\u57fa\u7840\u8bbe\u65bd\u5c42\u53ef\u80fd\u5b58\u5728\u7684\u95ee\u9898\u70b9\u5c31\u975e\u5e38\u591a\u4e86\u3002\u82e5\u9700\u8981\u5b9a\u4f4d\u4e00\u4e2a\u6bcf\u5c0f\u65f6\u4ec5\u5076\u53d1\u4e00\u6b21\u7684\u6296\u52a8\uff0c\u4f20\u7edf\u4e0a\u53ef\u80fd\u9700\u8981\u5728\u591a\u5904\u6293\u53d6tcpdump\u6216\u5206\u6790Nginx\u8bbf\u95ee\u65e5\u5fd7\uff0c\u6392\u67e5\u6548\u7387\u5f88\u4f4e\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic1.zhimg.com\/v2-bab2c0abc26514379a3a150fb46dc5a8_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">Metrics\uff1a\u5b9e\u7528eBPF\u91c7\u96c6LLM\u63a8\u7406\u670d\u52a1\u7684\u5168\u6808\u6307\u6807\u6027\u80fd<\/figcaption><\/figure>\n\n\n\n<p>DeepFlow\u5229\u7528eBPF\u91c7\u96c6\u6240\u6709\u7f51\u7edc\u8def\u5f84\u7684\u6d41\u91cf\u4e0e\u5ef6\u8fdf\u3002\u5e76\u5bf9\u5176\u65f6\u5ef6\u8fdb\u884c\u8ba1\u7b97\uff0c\u5c31\u80fd\u5f97\u5230\u4e00\u5f20\u5168\u6808\u8def\u5f84\u7684\u6307\u6807\u89c6\u56fe\u3002\u4ee5\u65f6\u5ef6\u4e3a\u4f8b\uff0c\u4ece\u5ba2\u6237\u7aef\u8fdb\u7a0b\u3001eBPF\u6355\u83b7\u7684\u7cfb\u7edf\u8c03\u7528\u3001\u5ba2\u6237\u7aefPod\u7684\u7f51\u5361\u3001\u5230K8s\u8282\u70b9\u7684\u7f51\u5361\uff0c\u5982\u679c\u8282\u70b9\u8fd8\u90e8\u7f72\u5728KVM\u5bbf\u4e3b\u673a\u4e0a\uff0c\u751a\u81f3\u53ef\u4ee5\u5ef6\u4f38\u5230\u5bbf\u4e3b\u673a\u8282\u70b9\u7684\u7f51\u5361\uff0c\u6574\u6761\u8def\u5f84\u4e0a\u6bcf\u4e00\u6bb5\u7684\u65f6\u5ef6\u6d88\u8017\u90fd\u80fd\u6e05\u6670\u5730\u5c55\u73b0\u51fa\u6765\u3002\u8fd9\u6837\u5c31\u80fd\u76f4\u63a5\u5224\u65ad\u95ee\u9898\u662f\u5426\u6e90\u4e8e\u57fa\u7840\u8bbe\u65bd\u5c42\uff0c\u907f\u514d\u4e86\u5728\u4e91\u539f\u751f\u73af\u5883\u4e2d\u642d\u5efaLLM\u670d\u52a1\u5e73\u53f0\u65f6\u5e38\u89c1\u7684\u5b9a\u754c\u4e0d\u6e05\u95ee\u9898\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>3.2 DeepFlow \u5168\u6808\u5168\u94fe\u8def\u8ffd\u8e2a\u89e3\u51b3\u65b9\u6848<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic1.zhimg.com\/v2-82a64a3903df15771fd5de86a0c37110_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">Tracing\uff1a\u4f7f\u7528eBPF\u5b9e\u73b0Python\u3001Golang\u63a8\u7406\u5f15\u64ce\u7684\u8ffd\u8e2a<\/figcaption><\/figure>\n\n\n\n<p>DeepFlow\u5229\u7528eBPF\u5b9e\u73b0\u4e86\u4e00\u79cd\u521b\u65b0\u6027\u7684AutoTracing\u80fd\u529b\uff0c\u80fd\u591f\u8986\u76d6\u57fa\u7840\u8bbe\u65bd\u7f51\u5173\u3001K8s\u7f51\u5361\u3001CoreDNS\u4e43\u81f3\u6570\u636e\u5e93\u7b49\u5168\u94fe\u8def\u7ec4\u4ef6\uff0c\u5e76\u80fd\u8f83\u597d\u5730\u8bc6\u522bPython\u7b49\u975eJava\u8bed\u8a00\u3002\u7279\u522b\u5730\uff0c\u4e3a\u5e94\u5bf9DeepSeek\u7b49\u6a21\u578b\u4e3a\u964d\u4f4e\u90e8\u7f72\u6210\u672c\u800c\u91c7\u7528\u7684\u78c1\u76d8KV Cache\u521b\u65b0\uff0cDeepFlow\u65b9\u6848\u4e5f\u80fd\u591f\u5c06\u76f8\u5173\u7684\u6587\u4ef6I\/O\u8c03\u7528\u4e8b\u4ef6\u5173\u8054\u81f3\u8ffd\u8e2a\u94fe\u8def\u4e2d\uff0c\u63d0\u4f9b\u5b58\u50a8\u6027\u80fd\u7684\u53ef\u89c2\u6d4b\u6027\u3002\u5982\u56fe\uff08Tracing\uff1a\u5b9e\u7528eBPF\u5b9e\u73b0Disk\/OSS KV Cache IO \u7684\u8ffd\u8e2a\uff09\u53f3\u4e0b\u89d2\u5219\u5c55\u793a\u4e86\u5982\u4f55\u5728DeepFlow Agent\u4e2d\u914d\u7f6e\uff0c\u4f7f\u5176\u5728\u91c7\u96c6\u4e00\u6b21\u8c03\u7528\u65f6\uff0c\u80fd\u5173\u8054\u91c7\u96c6\u5230\u76f8\u5e94\u7684IO\u4e8b\u4ef6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic3.zhimg.com\/v2-e157c987a4ebcbed6e0553f071f5c9a8_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">Tracing\uff1a\u5b9e\u7528eBPF\u5b9e\u73b0Disk\/OSS KV Cache IO \u7684\u8ffd\u8e2a<\/figcaption><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>DeepFlow \u5168\u6808\u51fd\u6570\u7ea7\u6027\u80fd\u5256\u6790\u89e3\u51b3\u65b9\u6848<\/strong><\/li>\n<\/ul>\n\n\n\n<p>eBPF \u80fd\u540c\u65f6\u6355\u83b7\u7528\u6237\u6001\uff08\u901a\u8fc7uprobe\uff09\u4e0e\u5185\u6838\u6001\u4e8b\u4ef6\uff0c\u5b9e\u73b0\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic1.zhimg.com\/v2-e879bd04afb8b0276be475f1734fdc92_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">Profiling\uff1a\u4f7f\u7528eBPF\u83b7\u53d6Python &amp; C++\u7684\u51fd\u6570\u8c03\u7528\u6808<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic1.zhimg.com\/v2-992436ae7199ee881bc9edaca2afbdd8_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">Profiling\uff1a\u4f7f\u7528eBPF\u83b7\u53d6Python &amp; C++\u7684\u51fd\u6570\u8c03\u7528\u6808<\/figcaption><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Python\u4e0eC++\u6df7\u5408\u8c03\u7528\u6808\u5408\u5e76<\/strong>\uff1a\u5c06vLLM\u7b49\u5f15\u64ce\u4e0a\u5c42\u7684Python\u51fd\u6570\u8c03\u7528\u4e0e\u5e95\u5c42\u7684C++\/CUDA\u5185\u6838\u8c03\u7528\u5173\u8054\uff0c\u5f62\u6210\u5b8c\u6574\u7684\u6027\u80fd\u706b\u7130\u56fe\u3002<\/li>\n\n\n\n<li><strong>\u7cbe\u51c6GPU\u8017\u65f6\u8ba1\u7b97<\/strong>\uff1a\u901a\u8fc7\u5206\u6790CUDA\u5185\u6838\u542f\u52a8\u4e0e\u540c\u6b65\u4e8b\u4ef6\uff0c\u7ed3\u5408CPU\u4fa7\u8c03\u7528\u6808\uff0c\u51c6\u786e\u8ba1\u7b97\u5f02\u6b65GPU\u64cd\u4f5c\u7684\u771f\u6b63\u8017\u65f6\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662fCPU\u7b49\u5f85\u65f6\u95f4\u3002<\/li>\n\n\n\n<li><strong>\u663e\u5b58\uff08HBM\uff09\u5256\u6790<\/strong>\uff1a\u53ef\u89c6\u5316\u663e\u5b58\u7533\u8bf7\u4e0e\u91ca\u653e\u7684\u70ed\u70b9\u51fd\u6570\uff0c\u8f85\u52a9\u5b9a\u4f4d\u5185\u5b58\u6ea2\u51fa\uff08OOM\uff09\u6839\u6e90\u3002\u5728\u8ba1\u7b97\u5bc6\u96c6\u578b\u573a\u666f\u4e2d\uff0cuprobe\u5e26\u6765\u7684\u6027\u80fd\u635f\u8017\u901a\u5e38\u4ec5\u4e3a1-5%\uff0c\u662f\u53ef\u63a5\u53d7\u7684\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic1.zhimg.com\/v2-804315b2a2b3a6cd417b4a0385944cfa_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">Profiling\uff1a\u8ba1\u7b97On-GPU Profiling<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001DeepFlow \u7528\u6237\u5b9e\u8df5<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6848\u4f8b\u4e00\uff1aDeepFlow\u667a\u80fd\u4f53\u7684\u81ea\u89c2\u6d4b\u5b9e\u8df5<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic2.zhimg.com\/v2-3b2d48ef3d9c1fdd3ba5500dad81ac9b_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">DeepFlow \u667a\u80fd\u4f53\uff1a\u4ea7\u54c1\u67b6\u6784<\/figcaption><\/figure>\n\n\n\n<p>DeepFlow\u5185\u90e8\u90e8\u7f72\u4e86\u4e00\u4e2a\u57fa\u4e8evLLM\u4e0eRay\u6784\u5efa\u7684AI\u667a\u80fd\u4f53\uff0c\u7528\u4e8e\u81ea\u52a8\u5316\u6545\u969c\u8bca\u65ad\u4e0e\u6570\u636e\u5206\u6790\u3002\u901a\u8fc7eBPF\u80fd\u529b\uff0c\u5b9e\u73b0\u4e86\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5bf9\u8be5\u667a\u80fd\u4f53\u63a8\u7406\u670d\u52a1\u7684GPU\u5185\u6838\u4e0e\u663e\u5b58\u64cd\u4f5c\u8fdb\u884c\u65e0\u4fb5\u5165\u5256\u6790\u3002<\/li>\n\n\n\n<li>\u6e05\u6670\u5c55\u793a\u6df7\u5408\u7f16\u7a0b\u6808\u4e0b\u7684\u6027\u80fd\u70ed\u70b9\uff0c\u6307\u5bfc\u5728\u4e0d\u540c\u578b\u53f7\u786c\u4ef6\u4e0a\u8fdb\u884c\u53c2\u6570\u8c03\u4f18\uff0c\u4ee5\u9002\u914d\u6700\u4f73\u670d\u52a1\u914d\u7f6e\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic3.zhimg.com\/v2-f09b405ba29f97353e0ad931cfe2572a_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">DeepFlow\u667a\u80fd\u4f53\uff1a\u4f7f\u7528Profiling\u5256\u6790vLLM+Ray\u63a8\u7406\u670d\u52a1<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic2.zhimg.com\/v2-7e432bfd124ce033b502f66534d45625_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">DeepFlow\u667a\u80fd\u4f53\uff1a\u4f7f\u7528Profiling\u5256\u6790vLLM+Ray\u63a8\u7406\u670d\u52a1<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pic2.zhimg.com\/v2-9e854a086feba7980da072ed906c1d95_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">DeepFlow\u667a\u80fd\u4f53\uff1a\u4f7f\u7528Profiling\u5256\u6790HBM\u7533\u8bf7\u548c\u4f7f\u7528<\/figcaption><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6848\u4f8b\u4e8c\uff1a\u4e2d\u56fd\u79fb\u52a8\u7684\u4e1a\u52a1\u6307\u6807\u91c7\u96c6<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/pica.zhimg.com\/v2-e4bcca921091e918b4a4bcba820c3424_1440w.jpg\" alt=\"\"\/><figcaption class=\"wp-element-caption\">\u4e2d\u56fd\u79fb\u52a8\uff1a\u6df1\u5ea6\u89e3\u6790DeepFlow\u5982\u4f55\u91c7\u96c6\u5927\u6a21\u578b\u670d\u52a1\u7684\u4e1a\u52a1\u6307\u6807<\/figcaption><\/figure>\n\n\n\n<p>\u4e2d\u56fd\u79fb\u52a8\u5229\u7528DeepFlow\u7684WASM\u63d2\u4ef6\u6269\u5c55\u80fd\u529b\uff0c\u5728\u4e0d\u4fee\u6539\u4e1a\u52a1\u4ee3\u7801\u7684\u60c5\u51b5\u4e0b\uff0c\u6210\u529f\u91c7\u96c6\u4e86\u5176\u79c1\u6709\u63a8\u7406\u534f\u8bae\u4e0b\u7684\u5168\u6808\u4e1a\u52a1\u6307\u6807\uff08\u5305\u62ecTTFT\u3001TPOT\u3001\u4ee4\u724c\u4ea7\u51fa\u7387\u7b49\uff09\uff0c\u5e76\u5728\u5fae\u670d\u52a1\u73af\u5883\u4e2d\u5b9e\u73b0\u4e86\u5168\u94fe\u8def\u6307\u6807\u6c47\u805a\u4e0e\u5206\u6790\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u7ed3\u8bba<\/h2>\n\n\n\n<p>\u57fa\u4e8eeBPF\u7684\u5168\u6808\u53ef\u89c2\u6d4b\u6027\u65b9\u6848\uff0c\u4e3a\u590d\u6742\u3001\u5f02\u6784\u7684LLM\u63a8\u7406\u670d\u52a1\u63d0\u4f9b\u4e86\u4ece\u5e94\u7528\u5230\u5e95\u5c42\u57fa\u7840\u8bbe\u65bd\u7684\u7edf\u4e00\u89c2\u6d4b\u89c6\u89d2\u3002\u5b83\u514b\u670d\u4e86\u4f20\u7edf\u63d2\u6869\u65b9\u6cd5\u7684\u4fb5\u5165\u6027\u5c40\u9650\u4e0e\u89c2\u6d4b\u6b7b\u89d2\uff0c\u901a\u8fc7\u5728\u7f51\u7edc\u3001\u7cfb\u7edf\u53ca\u5e94\u7528\u8fd0\u884c\u65f6\u5c42\u7684\u65e0\u7f1d\u5173\u8054\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6027\u80fd\u8bca\u65ad\u3001\u8d44\u6e90\u4f18\u5316\u4e0e\u6545\u969c\u5b9a\u4f4d\u7684\u6548\u7387\uff0c\u662f\u4e91\u539f\u751fAI\u57fa\u7840\u8bbe\u65bd\u53ef\u89c2\u6d4b\u6027\u6f14\u8fdb\u7684\u5173\u952e\u6280\u672f\u8def\u5f84\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p> \u672c\u6587\u7cfb\u7edf\u6027\u9610\u8ff0\u4e86\u5728\u4e91\u539f\u751f\u73af\u5883\u4e0b\uff0c\u5229\u7528\u6269\u5c55\u4f2f\u514b\u5229\u5305\u8fc7\u6ee4\u5668\uff08eBPF\uff09\u6280\u672f\u5b9e\u73b0\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u63a8\u7406\u670d\u52a1\u5168\u6808\u53ef\u89c2\u6d4b\u6027\u7684\u65b9\u6cd5\u4e0e\u67b6\u6784\u3002\u5185\u5bb9\u6db5\u76d6\u81ea\u5efa\u5343\u4ebf\u53c2\u6570LLM\u63a8\u7406\u670d\u52a1\u5728\u5f02\u6784\u786c\u4ef6\uff08\u5982\u6607\u817e910B\uff09\u4e0e\u5206\u5e03\u5f0f\u6846\u67b6\uff08\u5982vLLM, Ray\uff09\u4e0b\u9762\u4e34\u7684\u6311\u6218\uff0c\u5206\u6790\u4e86\u4f20\u7edf\u53ef\u89c2\u6d4b\u6027\u4e09\u5927\u652f\u67f1\uff08\u6307\u6807\u3001\u8ffd\u8e2a\u3001\u65e5\u5fd7\uff09\u53ca\u65b0\u5174\u6027\u80fd\u5256\u6790\u652f\u67f1\u7684\u4e0d\u8db3\u3002\u6587\u7ae0\u91cd\u70b9\u4ecb\u7ecd\u4e86DeepFlow\u5e73\u53f0\u5982\u4f55\u901a\u8fc7eBPF\u5b9e\u73b0\u96f6\u4ee3\u7801\u4fb5\u5165\u7684\u5168\u6808\u6307\u6807\u91c7\u96c6\u3001\u5168\u94fe\u8def\u8ffd\u8e2a\u4e0e\u6df7\u5408\u6808\uff08CPU\/GPU\uff09\u6027\u80fd\u5256\u6790\uff0c\u5e76\u8f85\u4ee5\u667a\u80fd\u4f53\u5e94\u7528\u4e0e\u4e2d\u56fd\u79fb\u52a8\u7684\u843d\u5730\u6848\u4f8b\uff0c\u9a8c\u8bc1\u4e86\u8be5\u65b9\u6848\u5728\u4f18\u5316GPU\u5229\u7528\u7387\u3001\u5b9a\u4f4d\u63a8\u7406\u5ef6\u8fdf\u53ca\u9002\u914d\u6d41\u5f0f\u534f\u8bae\u7b49\u65b9\u9762\u7684\u6709\u6548\u6027\u3002<\/p>\n","protected":false},"author":1,"featured_media":42207,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12,6,7,11],"tags":[235],"class_list":["post-42205","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ys-news","category-news","category-top","category-solution","tag-ebpfllmgpudeepflow-2"],"_links":{"self":[{"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/posts\/42205","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/comments?post=42205"}],"version-history":[{"count":2,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/posts\/42205\/revisions"}],"predecessor-version":[{"id":42221,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/posts\/42205\/revisions\/42221"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/media\/42207"}],"wp:attachment":[{"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/media?parent=42205"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/categories?post=42205"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/tags?post=42205"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}