{"id":33376,"date":"2025-03-04T12:04:11","date_gmt":"2025-03-04T04:04:11","guid":{"rendered":"https:\/\/www.yunshan.net\/blog\/?p=33376"},"modified":"2025-03-04T12:04:11","modified_gmt":"2025-03-04T04:04:11","slug":"ebpf%e5%a6%82%e4%bd%95%e5%ae%9e%e7%8e%b0%e5%af%b9gpu%e8%b5%84%e6%ba%90%e7%9a%84%e7%9b%91%e6%8e%a7%e4%b8%8e%e4%bc%98%e5%8c%96%ef%bc%9f","status":"publish","type":"post","link":"https:\/\/www.yunshan.net\/blog\/archives\/33376.html","title":{"rendered":"eBPF\u5982\u4f55\u5b9e\u73b0\u5bf9GPU\u8d44\u6e90\u7684\u76d1\u63a7\u4e0e\u4f18\u5316\uff1f"},"content":{"rendered":"<p style=\"text-align: center;\">\n<p><img decoding=\"async\" src=\"https:\/\/www.yunshan.net\/blog\/wp-content\/uploads\/2025\/03\/4wdpi5AT.jpg\" alt=\"eBPF\u5982\u4f55\u5b9e\u73b0\u5bf9GPU\u8d44\u6e90\u7684\u76d1\u63a7\u4e0e\u4f18\u5316\uff1f\" ><\/p>\n<\/p>\n<hr>\n<p><strong>\u524d\u8a00<\/strong><br \/>\n\u5728\u4eba\u5de5\u667a\u80fd\u3001\u6df1\u5ea6\u5b66\u4e60\u4e0e\u56fe\u5f62\u6e32\u67d3\u9886\u57df\uff0cGPU\u7684\u8ba1\u7b97\u80fd\u529b\u5df2\u6210\u4e3a\u9a71\u52a8\u6280\u672f\u9769\u65b0\u7684\u6838\u5fc3\u5f15\u64ce\u3002\u7136\u800c\uff0c\u968f\u7740\u7b97\u529b\u9700\u6c42\u7684\u6307\u6570\u7ea7\u589e\u957f\uff0c\u5982\u4f55\u9ad8\u6548\u76d1\u63a7\u4e0e\u4f18\u5316GPU\u8d44\u6e90\uff0c\u6210\u4e3a\u5f00\u53d1\u8005\u4e0e\u8fd0\u7ef4\u56e2\u961f\u9762\u4e34\u7684\u5173\u952e\u6311\u6218\u3002\u4f20\u7edf\u5de5\u5177\u5f80\u5f80\u96be\u4ee5\u7a7f\u900f\u786c\u4ef6\u62bd\u8c61\u5c42\uff0c\u5bfc\u81f4\u8d44\u6e90\u5229\u7528\u7387\u4f4e\u4e0b\u6216\u6027\u80fd\u74f6\u9888\u96be\u4ee5\u5b9a\u4f4d\u3002\u6b64\u65f6\uff0c<strong><a href=\"https:\/\/www.yunshan.net\/\"   target=\"_blank\" style=\"color:blue\">eBPF<\/a>\uff08Extended Berkeley Packet Filter\uff09<\/strong>\u51ed\u501f\u5176\u5185\u6838\u7ea7\u53ef\u89c2\u6d4b\u6027\u4e0e\u52a8\u6001\u6ce8\u5165\u80fd\u529b\uff0c\u4e3aGPU\u8d44\u6e90\u7684\u7cbe\u7ec6\u5316\u7ba1\u63a7\u63d0\u4f9b\u4e86\u5168\u65b0\u601d\u8def\u3002\u672c\u6587\u5c06\u6df1\u5165\u63a2\u8ba8<a href=\"https:\/\/www.yunshan.net\/\"   target=\"_blank\" style=\"color:blue\">eBPF<\/a>\u5982\u4f55\u7a81\u7834\u4f20\u7edf\u76d1\u63a7\u5de5\u5177\u7684\u5c40\u9650\uff0c\u5b9e\u73b0\u4ece\u5185\u6838\u6001\u5230\u7528\u6237\u6001\u7684\u7aef\u5230\u7aefGPU\u6027\u80fd\u6d1e\u5bdf\uff0c\u5e76\u901a\u8fc7\u771f\u5b9e\u6848\u4f8b\u63ed\u793a\u5176\u5728\u8d44\u6e90\u4f18\u5316\u4e2d\u7684\u72ec\u7279\u4ef7\u503c\u3002<\/p>\n<hr>\n<h3 id=\"\u4e00-ebpf-\u91cd\u65b0\u5b9a\u4e49\u7cfb\u7edf\u53ef\u89c2\u6d4b\u6027\"><strong>\u4e00\u3001eBPF\uff1a\u91cd\u65b0\u5b9a\u4e49\u7cfb\u7edf\u53ef\u89c2\u6d4b\u6027<\/strong><\/h3>\n<p>eBPF\u6700\u521d\u88ab\u8bbe\u8ba1\u4e3a\u7f51\u7edc\u6570\u636e\u5305\u8fc7\u6ee4\u5de5\u5177\uff0c\u4f46\u7ecf\u8fc7\u591a\u5e74\u53d1\u5c55\uff0c\u5176\u80fd\u529b\u5df2\u6269\u5c55\u81f3\u7cfb\u7edf\u8ddf\u8e2a\u3001\u6027\u80fd\u5206\u6790\u53ca\u5b89\u5168\u76d1\u63a7\u7b49\u9886\u57df\u3002<strong>\u901a\u8fc7\u5728\u5185\u6838\u4e2d\u8fd0\u884c\u6c99\u76d2\u5316\u7a0b\u5e8f\uff0ceBPF\u65e0\u9700\u4fee\u6539\u5185\u6838\u4ee3\u7801\u5373\u53ef\u52a8\u6001\u6ce8\u5165\u63a2\u9488<\/strong>\uff0c\u5b9e\u65f6\u6355\u83b7\u7cfb\u7edf\u8c03\u7528\u3001\u51fd\u6570\u6267\u884c\u751a\u81f3\u786c\u4ef6\u4e8b\u4ef6\u3002\u8fd9\u79cd\u7279\u6027\u4f7f\u5176\u80fd\u591f\u4ee5\u6781\u4f4e\u5f00\u9500\u83b7\u53d6\u4f20\u7edf\u5de5\u5177\u96be\u4ee5\u89e6\u53ca\u7684\u6df1\u5c42\u6570\u636e\uff0c\u4f8b\u5982GPU\u9a71\u52a8\u4e0e\u5e94\u7528\u7a0b\u5e8f\u95f4\u7684\u4ea4\u4e92\u7ec6\u8282\u3002<\/p>\n<p>\u4e0e\u57fa\u4e8e\u91c7\u6837\uff08\u5982<code>perf<\/code>\uff09\u6216\u65e5\u5fd7\u5206\u6790\u7684\u4f20\u7edf\u76d1\u63a7\u65b9\u5f0f\u4e0d\u540c\uff0ceBPF\u652f\u6301<strong>\u4e8b\u4ef6\u9a71\u52a8\u7684\u8fde\u7eed\u8ffd\u8e2a<\/strong>\u3002\u4f8b\u5982\uff0c\u5f53GPU\u5185\u5b58\u5206\u914d\u8d85\u8fc7\u9608\u503c\u65f6\uff0ceBPF\u7a0b\u5e8f\u53ef\u7acb\u5373\u89e6\u53d1\u544a\u8b66\uff0c\u800c\u975e\u7b49\u5f85\u5468\u671f\u6027\u91c7\u6837\u3002\u8fd9\u4e00\u80fd\u529b\u5bf9GPU\u8fd9\u7c7b\u9ad8\u541e\u5410\u3001\u4f4e\u5ef6\u8fdf\u8bbe\u5907\u5c24\u4e3a\u91cd\u8981\u3002<\/p>\n<hr>\n<h3 id=\"\u4e8c-gpu\u76d1\u63a7\u7684\u6311\u6218\u4e0eebpf\u7684\u7834\u5c40\u4e4b\u9053\"><strong>\u4e8c\u3001GPU\u76d1\u63a7\u7684\u6311\u6218\u4e0eeBPF\u7684\u7834\u5c40\u4e4b\u9053<\/strong><\/h3>\n<h4 id=\"1-gpu\u8d44\u6e90\u76d1\u63a7\u7684\u75db\u70b9\"><strong>1. GPU\u8d44\u6e90\u76d1\u63a7\u7684\u75db\u70b9<\/strong><\/h4>\n<p>GPU\u901a\u5e38\u901a\u8fc7\u4e13\u6709\u9a71\u52a8\uff08\u5982NVIDIA CUDA\uff09\u4e0e\u4e0a\u5c42\u5e94\u7528\u4ea4\u4e92\uff0c\u5176\u5185\u90e8\u72b6\u6001\uff08\u5982SM\u5360\u7528\u7387\u3001\u663e\u5b58\u788e\u7247\u3001\u6838\u51fd\u6570\u6267\u884c\u65f6\u957f\uff09\u5bf9\u7528\u6237\u6001\u5de5\u5177\u5e76\u4e0d\u900f\u660e\u3002\u4f20\u7edf\u76d1\u63a7\u624b\u6bb5\u4f9d\u8d56\u5382\u5546\u63d0\u4f9b\u7684\u63a5\u53e3\uff08\u5982<code>nvidia-smi<\/code>\uff09\uff0c\u5b58\u5728\u4e24\u5927\u7f3a\u9677\uff1a<\/p>\n<ul>\n<li><strong>\u6570\u636e\u7c92\u5ea6\u7c97\u7cd9<\/strong>\uff1a\u4ec5\u80fd\u83b7\u53d6\u5168\u5c40\u7ea7\u522b\u7684\u5229\u7528\u7387\u7edf\u8ba1\uff0c\u65e0\u6cd5\u5173\u8054\u5230\u5177\u4f53\u8fdb\u7a0b\u6216\u5bb9\u5668\uff1b\n<\/li>\n<li><strong>\u5ef6\u8fdf\u9ad8<\/strong>\uff1a\u91c7\u6837\u95f4\u9694\u901a\u5e38\u4e3a\u79d2\u7ea7\uff0c\u96be\u4ee5\u6355\u6349\u77ac\u65f6\u74f6\u9888\uff08\u5982\u5fae\u79d2\u7ea7\u6838\u51fd\u6570\u963b\u585e\uff09\u3002\n<\/li>\n<\/ul>\n<h4 id=\"2-ebpf\u7684\u89e3\u51b3\u65b9\u6848\"><strong>2. eBPF\u7684\u89e3\u51b3\u65b9\u6848<\/strong><\/h4>\n<p>eBPF\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u7a81\u7834\u4e0a\u8ff0\u9650\u5236\uff1a<\/p>\n<ul>\n<li><strong>\u52a8\u6001\u63d2\u6869GPU\u9a71\u52a8\u51fd\u6570<\/strong>\uff1a\u5728\u5185\u6838\u6001\u62e6\u622aCUDA\u9a71\u52a8\u4e0e\u786c\u4ef6\u7684\u4ea4\u4e92\uff08\u5982<code>cuMemAlloc<\/code>\u3001<code>cuLaunchKernel<\/code>\uff09\uff0c\u8bb0\u5f55\u6bcf\u6b21\u8c03\u7528\u7684\u65f6\u95f4\u6233\u3001\u8fdb\u7a0bID\u53ca\u53c2\u6570\uff1b\n<\/li>\n<li><strong>\u5173\u8054\u7528\u6237\u6001\u4e0e\u5185\u6838\u6001\u4e8b\u4ef6<\/strong>\uff1a\u7ed3\u5408<code>uprobe<\/code>\u8ffd\u8e2a\u5e94\u7528\u7a0b\u5e8f\u7684CUDA API\u8c03\u7528\uff08\u5982<code>cudaMalloc<\/code>\uff09\uff0c\u5f62\u6210\u4ece\u5e94\u7528\u4ee3\u7801\u5230\u786c\u4ef6\u6267\u884c\u7684\u5b8c\u6574\u8c03\u7528\u94fe\uff1b\n<\/li>\n<li><strong>\u4f4e\u5f00\u9500\u6570\u636e\u805a\u5408<\/strong>\uff1a\u5728\u5185\u6838\u4e2d\u76f4\u63a5\u8fc7\u6ee4\u3001\u7edf\u8ba1\u4e8b\u4ef6\uff0c\u4ec5\u5c06\u7ed3\u679c\u4f20\u9012\u7ed9\u7528\u6237\u6001\uff0c\u907f\u514d\u6d77\u91cf\u539f\u59cb\u6570\u636e\u590d\u5236\u9020\u6210\u7684\u6027\u80fd\u635f\u8017\u3002\n<\/li>\n<\/ul>\n<p><strong>\u901a\u8fc7\u6302\u94a9NVIDIA\u9a71\u52a8\u4e2d\u7684\u5185\u5b58\u5206\u914d\u51fd\u6570\uff0ceBPF\u53ef\u5b9e\u65f6\u7edf\u8ba1\u6bcf\u4e2a\u5bb9\u5668\u7684\u663e\u5b58\u5360\u7528\uff0c\u5e76\u5728\u8d85\u9650\u65f6\u89e6\u53d1\u56de\u6536\u7b56\u7565<\/strong>\uff0c\u907f\u514d\u56e0\u5355\u5bb9\u5668\u8d44\u6e90\u6ee5\u7528\u5bfc\u81f4\u7684\u6574\u4f53\u6027\u80fd\u4e0b\u964d\u3002<\/p>\n<hr>\n<h3 id=\"\u4e09-ebpf\u76d1\u63a7gpu\u7684\u6838\u5fc3\u6280\u672f\u8def\u5f84\"><strong>\u4e09\u3001eBPF\u76d1\u63a7GPU\u7684\u6838\u5fc3\u6280\u672f\u8def\u5f84<\/strong><\/h3>\n<h4 id=\"1-\u5185\u6838\u6001\u63a2\u9488\u6ce8\u5165\"><strong>1. \u5185\u6838\u6001\u63a2\u9488\u6ce8\u5165<\/strong><\/h4>\n<p>eBPF\u7a0b\u5e8f\u901a\u8fc7<code>kprobe<\/code>\u6216<code>tracepoint<\/code>\u6302\u8f7d\u5230GPU\u9a71\u52a8\u7684\u5173\u952e\u51fd\u6570\u3002\u4ee5\u76d1\u63a7CUDA\u6838\u51fd\u6570\u6267\u884c\u4e3a\u4f8b\uff1a<\/p>\n<pre><code class=\"language-c\">SEC(&quot;kprobe\/cuLaunchKernel&quot;)  \r\nint trace_cu_launch(struct pt_regs *ctx) {  \r\nu64 pid = bpf_get_current_pid_tgid();  \r\nu64 duration_ns = bpf_ktime_get_ns();  \r\n\/\/ \u8bb0\u5f55\u8fdb\u7a0bPID\u3001\u6838\u51fd\u6570\u540d\u79f0\u53ca\u542f\u52a8\u65f6\u95f4  \r\nbpf_map_update_elem(&amp;kernel_events, &amp;pid, &amp;duration_ns, BPF_ANY);  \r\nreturn 0;  \r\n}  \r\n<\/code><\/pre>\n<p>\u6b64\u7c7b\u7a0b\u5e8f\u53ef\u7cbe\u786e\u7edf\u8ba1\u6bcf\u4e2a\u6838\u51fd\u6570\u7684\u6267\u884c\u65f6\u957f\uff0c\u5e76\u5173\u8054\u5230\u53d1\u8d77\u8fdb\u7a0b\uff0c\u8fdb\u800c\u8bc6\u522b\u51fa\u4f4e\u6548\u7684GPU\u4ee3\u7801\u903b\u8f91\u3002<\/p>\n<h4 id=\"2-\u7528\u6237\u6001-\u5185\u6838\u6001\u6570\u636e\u534f\u540c\"><strong>2. \u7528\u6237\u6001-\u5185\u6838\u6001\u6570\u636e\u534f\u540c<\/strong><\/h4>\n<p>eBPF Maps\u4f5c\u4e3a\u5185\u6838\u4e0e\u7528\u6237\u6001\u5171\u4eab\u7684\u6570\u636e\u7ed3\u6784\uff0c\u652f\u6301\u5b9e\u65f6\u4f20\u9012\u76d1\u63a7\u6307\u6807\u3002\u4f8b\u5982\uff0c<strong>\u901a\u8fc7Hash Map\u7edf\u8ba1\u5404\u8fdb\u7a0b\u7684\u663e\u5b58\u5206\u914d\u603b\u91cf<\/strong>\uff0c\u7528\u6237\u6001\u5de5\u5177\u53ef\u6bcf\u79d2\u8bfb\u53d6\u4e00\u6b21\uff0c\u751f\u6210\u52a8\u6001\u62d3\u6251\u56fe\u3002\u8fd9\u79cd\u8bbe\u8ba1\u907f\u514d\u4e86\u9891\u7e41\u4e0a\u4e0b\u6587\u5207\u6362\uff0c\u786e\u4fdd\u76d1\u63a7\u5f00\u9500\u4f4e\u4e8e1% CPU\u5360\u7528\u3002<\/p>\n<h4 id=\"3-\u4e0e\u5bb9\u5668\u5316\u73af\u5883\u7684\u96c6\u6210\"><strong>3. \u4e0e\u5bb9\u5668\u5316\u73af\u5883\u7684\u96c6\u6210<\/strong><\/h4>\n<p>\u5728Kubernetes\u96c6\u7fa4\u4e2d\uff0ceBPF\u53ef\u7ed3\u5408<code>cgroup<\/code>\u4fe1\u606f\uff0c\u5c06GPU\u6307\u6807\u5173\u8054\u5230Pod\u6216\u547d\u540d\u7a7a\u95f4\u3002\u5f00\u6e90\u9879\u76ee<strong>Kubectl-GPU<\/strong>\u5373\u5229\u7528eBPF\u5b9e\u73b0\u4e86\u5bb9\u5668\u7ea7\u522b\u7684GPU\u5229\u7528\u7387\u76d1\u63a7\uff0c\u5e2e\u52a9\u8c03\u5ea6\u5668\u4f18\u5316\u8d44\u6e90\u5206\u914d\u7b56\u7565\u3002<\/p>\n<hr>\n<h3 id=\"\u56db-\u4ece\u76d1\u63a7\u5230\u4f18\u5316-ebpf\u7684\u5b9e\u8df5\u573a\u666f\"><strong>\u56db\u3001\u4ece\u76d1\u63a7\u5230\u4f18\u5316\uff1aeBPF\u7684\u5b9e\u8df5\u573a\u666f<\/strong><\/h3>\n<h4 id=\"1-\u663e\u5b58\u6cc4\u6f0f\u5b9a\u4f4d\"><strong>1. \u663e\u5b58\u6cc4\u6f0f\u5b9a\u4f4d<\/strong><\/h4>\n<p>\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u4efb\u52a1\u5e38\u56e0\u672a\u91ca\u653e\u4e34\u65f6\u5f20\u91cf\u5bfc\u81f4\u663e\u5b58\u6cc4\u6f0f\u3002<strong>\u901a\u8fc7\u6302\u94a9<code>cuMemAlloc<\/code>\u4e0e<code>cuMemFree<\/code>\uff0ceBPF\u53ef\u6784\u5efa\u663e\u5b58\u5206\u914d\u7684\u751f\u547d\u5468\u671f\u56fe\u8c31<\/strong>\uff0c\u5e76\u81ea\u52a8\u8bc6\u522b\u672a\u914d\u5bf9\u7684\u5206\u914d\u64cd\u4f5c\u3002\u67d0AI\u5e73\u53f0\u4f7f\u7528\u6b64\u65b9\u6cd5\u5c06\u6cc4\u6f0f\u6392\u67e5\u65f6\u95f4\u4ece\u6570\u5c0f\u65f6\u7f29\u77ed\u81f3\u5206\u949f\u7ea7\u3002<\/p>\n<h4 id=\"2-\u6838\u51fd\u6570\u6027\u80fd\u5206\u6790\"><strong>2. \u6838\u51fd\u6570\u6027\u80fd\u5206\u6790<\/strong><\/h4>\n<p>eBPF\u53ef\u7edf\u8ba1\u6bcf\u4e2aCUDA\u6838\u51fd\u6570\u7684\u6267\u884c\u65f6\u957f\u53ca\u5176\u8c03\u7528\u4e0a\u4e0b\u6587\u3002\u67d0\u81ea\u52a8\u9a7e\u9a76\u56e2\u961f\u53d1\u73b0\uff0c\u5176\u9884\u5904\u7406\u6d41\u6c34\u7ebf\u4e2d\u67d0\u4e2a\u6838\u51fd\u6570\u56e0\u5bc4\u5b58\u5668\u6ea2\u51fa\u5bfc\u81f4\u8017\u65f6\u5f02\u5e38\uff0c<strong>\u901a\u8fc7eBPF\u751f\u6210\u7684\u706b\u7130\u56fe\u5feb\u901f\u5b9a\u4f4d\u95ee\u9898\uff0c\u4f18\u5316\u540e\u63a8\u7406\u5ef6\u8fdf\u964d\u4f4e18%<\/strong>\u3002<\/p>\n<h4 id=\"3-\u591a\u79df\u6237\u8d44\u6e90\u9694\u79bb\"><strong>3. \u591a\u79df\u6237\u8d44\u6e90\u9694\u79bb<\/strong><\/h4>\n<p>\u5728\u4e91\u73af\u5883\u4e2d\uff0ceBPF\u53ef\u5f3a\u5236\u5b9e\u65bdGPU\u663e\u5b58\u4e0e\u7b97\u529b\u914d\u989d\u3002\u4f8b\u5982\uff0c<strong>\u57fa\u4e8eeBPF\u7684QoS\u63a7\u5236\u5668\u53ef\u5728\u9a71\u52a8\u5c42\u62e6\u622a\u8d85\u9650\u8bf7\u6c42<\/strong>\uff0c\u786e\u4fdd\u9ad8\u4f18\u5148\u7ea7\u4efb\u52a1\u4e0d\u53d7\u201c\u566a\u58f0\u90bb\u5c45\u201d\u5f71\u54cd\u3002\u5b9e\u6d4b\u663e\u793a\uff0c\u8be5\u65b9\u6cd5\u53ef\u5c06GPU\u5229\u7528\u7387\u6ce2\u52a8\u8303\u56f4\u4ece\u00b140%\u538b\u7f29\u81f3\u00b15%\u3002<\/p>\n<h4 id=\"4-\u80fd\u8017\u4f18\u5316\"><strong>4. \u80fd\u8017\u4f18\u5316<\/strong><\/h4>\n<p>\u901a\u8fc7\u5173\u8054GPU\u5229\u7528\u7387\u4e0e\u529f\u8017\u4f20\u611f\u5668\u6570\u636e\uff0ceBPF\u53ef\u5efa\u7acb\u529f\u8017\u6a21\u578b\uff0c\u8bc6\u522b\u9ad8\u80fd\u8017\u4f4e\u6548\u4efb\u52a1\u3002\u67d0\u8d85\u7b97\u4e2d\u5fc3\u5229\u7528\u6b64\u6a21\u578b\u8c03\u6574\u4f5c\u4e1a\u8c03\u5ea6\u7b56\u7565\uff0c\u5168\u5e74\u8282\u7535\u8d85120\u4e07\u5343\u74e6\u65f6\u3002<\/p>\n<hr>\n<h3 id=\"\u4e94-\u6848\u4f8b-ebpf\u5728ai\u8bad\u7ec3\u96c6\u7fa4\u4e2d\u7684\u843d\u5730\"><strong>\u4e94\u3001\u6848\u4f8b\uff1aeBPF\u5728AI\u8bad\u7ec3\u96c6\u7fa4\u4e2d\u7684\u843d\u5730<\/strong><\/h3>\n<p>\u67d0\u5934\u90e8\u7535\u5546\u7684\u63a8\u8350\u7cfb\u7edf\u9700\u540c\u65f6\u8fd0\u884c\u6570\u767e\u4e2aTensorFlow\u8bad\u7ec3\u4efb\u52a1\u3002\u7531\u4e8e\u7f3a\u4e4f\u7ec6\u7c92\u5ea6\u76d1\u63a7\uff0c\u96c6\u7fa4\u5e38\u56e0\u663e\u5b58\u7ade\u4e89\u5bfc\u81f4\u4efb\u52a1\u5931\u8d25\u3002\u90e8\u7f72\u57fa\u4e8eeBPF\u7684\u76d1\u63a7\u7cfb\u7edf\u540e\uff0c\u8fd0\u7ef4\u56e2\u961f\u5b9e\u73b0\u4e86\u4ee5\u4e0b\u6539\u8fdb\uff1a<\/p>\n<ul>\n<li><strong>\u5b9e\u65f6\u663e\u5b58\u70ed\u5ea6\u56fe<\/strong>\uff1a\u6309\u5bb9\u5668\u7ef4\u5ea6\u5c55\u793a\u663e\u5b58\u5360\u7528\uff0c\u5feb\u901f\u8bc6\u522b\u8d44\u6e90\u4e89\u7528\uff1b\n<\/li>\n<li><strong>\u81ea\u52a8\u5f02\u5e38\u68c0\u6d4b<\/strong>\uff1a\u5f53\u6838\u51fd\u6570\u6267\u884c\u65f6\u95f4\u8d85\u8fc7\u5386\u53f2\u57fa\u7ebf3\u500d\u65f6\u89e6\u53d1\u544a\u8b66\uff1b\n<\/li>\n<li><strong>\u52a8\u6001\u4f18\u5148\u7ea7\u8c03\u6574<\/strong>\uff1a\u7ed3\u5408eBPF\u6307\u6807\uff0c\u8c03\u5ea6\u5668\u4e3a\u5173\u952e\u4efb\u52a1\u9884\u7559\u5e26\u5bbd\u3002\n<\/li>\n<\/ul>\n<p><strong>\u7ed3\u679c<\/strong>\uff1a\u4efb\u52a1\u5931\u8d25\u7387\u4e0b\u964d67%\uff0c\u5e73\u5747\u8bad\u7ec3\u5468\u671f\u7f29\u77ed23%\uff0c\u5e74\u5ea6\u786c\u4ef6\u6210\u672c\u8282\u7701\u7ea6280\u4e07\u7f8e\u5143\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u524d\u8a00 \u5728\u4eba\u5de5\u667a\u80fd\u3001\u6df1\u5ea6\u5b66\u4e60\u4e0e\u56fe\u5f62\u6e32\u67d3\u9886\u57df\uff0cGPU\u7684\u8ba1\u7b97\u80fd\u529b\u5df2\u6210\u4e3a\u9a71\u52a8\u6280\u672f\u9769\u65b0\u7684\u6838\u5fc3\u5f15\u64ce\u3002\u7136\u800c\uff0c\u968f\u7740\u7b97\u529b\u9700\u6c42\u7684\u6307\u6570\u7ea7\u589e\u957f\uff0c\u5982\u4f55\u9ad8\u6548\u76d1\u63a7\u4e0e\u4f18\u5316GPU\u8d44\u6e90\uff0c\u6210\u4e3a\u5f00\u53d1\u8005\u4e0e\u8fd0\u7ef4\u56e2\u961f\u9762\u4e34\u7684\u5173\u952e\u6311\u6218\u3002\u4f20\u7edf\u5de5\u5177\u5f80\u5f80\u96be\u4ee5\u7a7f\u900f\u786c\u4ef6\u62bd\u8c61\u5c42\uff0c\u5bfc\u81f4\u8d44\u6e90\u5229\u7528\u7387\u4f4e\u4e0b\u6216\u6027\u80fd\u74f6\u9888\u96be\u4ee5\u5b9a\u4f4d\u3002\u6b64\u65f6\uff0ceBPF\uff08Extended Berkeley Packet Filter\uff09\u51ed\u501f\u5176\u5185\u6838\u7ea7\u53ef\u89c2\u6d4b\u6027\u4e0e\u52a8\u6001\u6ce8\u5165\u80fd\u529b\uff0c\u4e3aGPU\u8d44\u6e90\u7684\u7cbe\u7ec6\u5316\u7ba1\u63a7\u63d0\u4f9b\u4e86\u5168\u65b0\u601d\u8def\u3002\u672c\u6587\u5c06\u6df1\u5165\u63a2\u8ba8eBPF\u5982\u4f55\u7a81\u7834\u4f20\u7edf\u76d1\u63a7\u5de5\u5177\u7684\u5c40\u9650\uff0c\u5b9e\u73b0\u4ece\u5185\u6838\u6001\u5230\u7528\u6237\u6001\u7684\u7aef\u5230\u7aefGPU\u6027\u80fd\u6d1e\u5bdf\uff0c\u5e76\u901a\u8fc7\u771f\u5b9e\u6848\u4f8b\u63ed\u793a\u5176\u5728\u8d44\u6e90\u4f18\u5316\u4e2d\u7684\u72ec\u7279\u4ef7\u503c\u3002 \u4e00\u3001eBPF\uff1a\u91cd\u65b0\u5b9a\u4e49\u7cfb\u7edf\u53ef\u89c2\u6d4b\u6027 eBPF\u6700\u521d\u88ab\u8bbe\u8ba1\u4e3a\u7f51\u7edc\u6570\u636e\u5305\u8fc7\u6ee4\u5de5\u5177\uff0c\u4f46\u7ecf\u8fc7\u591a\u5e74\u53d1\u5c55\uff0c\u5176\u80fd\u529b\u5df2\u6269\u5c55\u81f3\u7cfb\u7edf\u8ddf\u8e2a\u3001\u6027\u80fd\u5206\u6790\u53ca\u5b89\u5168\u76d1\u63a7\u7b49\u9886\u57df\u3002\u901a\u8fc7\u5728\u5185\u6838\u4e2d\u8fd0\u884c\u6c99\u76d2\u5316\u7a0b\u5e8f\uff0ceBPF\u65e0\u9700\u4fee\u6539\u5185\u6838\u4ee3\u7801\u5373\u53ef\u52a8\u6001\u6ce8\u5165\u63a2\u9488\uff0c\u5b9e\u65f6\u6355\u83b7\u7cfb\u7edf\u8c03\u7528\u3001\u51fd\u6570\u6267\u884c\u751a\u81f3\u786c\u4ef6\u4e8b\u4ef6\u3002\u8fd9\u79cd\u7279\u6027\u4f7f\u5176\u80fd\u591f\u4ee5\u6781\u4f4e\u5f00\u9500\u83b7\u53d6\u4f20\u7edf\u5de5\u5177\u96be\u4ee5\u89e6\u53ca\u7684\u6df1\u5c42\u6570\u636e\uff0c\u4f8b\u5982GPU\u9a71\u52a8\u4e0e\u5e94\u7528\u7a0b\u5e8f\u95f4\u7684\u4ea4\u4e92\u7ec6\u8282\u3002 \u4e0e\u57fa\u4e8e\u91c7\u6837\uff08\u5982perf\uff09\u6216\u65e5\u5fd7\u5206\u6790\u7684\u4f20\u7edf\u76d1\u63a7\u65b9\u5f0f\u4e0d\u540c\uff0ceBPF\u652f\u6301\u4e8b\u4ef6\u9a71\u52a8\u7684\u8fde\u7eed\u8ffd\u8e2a\u3002\u4f8b\u5982\uff0c\u5f53GPU\u5185\u5b58\u5206\u914d\u8d85\u8fc7\u9608\u503c\u65f6\uff0ceBPF\u7a0b\u5e8f\u53ef\u7acb\u5373\u89e6\u53d1\u544a\u8b66\uff0c\u800c\u975e\u7b49\u5f85\u5468\u671f\u6027\u91c7\u6837\u3002\u8fd9\u4e00\u80fd\u529b\u5bf9GPU\u8fd9\u7c7b\u9ad8\u541e\u5410\u3001\u4f4e\u5ef6\u8fdf\u8bbe\u5907\u5c24\u4e3a\u91cd\u8981\u3002 \u4e8c\u3001GPU\u76d1\u63a7\u7684\u6311\u6218\u4e0eeBPF\u7684\u7834\u5c40\u4e4b\u9053 1. GPU\u8d44\u6e90\u76d1\u63a7\u7684\u75db\u70b9 GPU\u901a\u5e38\u901a\u8fc7\u4e13\u6709\u9a71\u52a8\uff08\u5982NVIDIA CUDA\uff09\u4e0e\u4e0a\u5c42\u5e94\u7528\u4ea4\u4e92\uff0c\u5176\u5185\u90e8\u72b6\u6001\uff08\u5982SM\u5360\u7528\u7387\u3001\u663e\u5b58\u788e\u7247\u3001\u6838\u51fd\u6570\u6267\u884c\u65f6\u957f\uff09\u5bf9\u7528\u6237\u6001\u5de5\u5177\u5e76\u4e0d\u900f\u660e\u3002\u4f20\u7edf\u76d1\u63a7\u624b\u6bb5\u4f9d\u8d56\u5382\u5546\u63d0\u4f9b\u7684\u63a5\u53e3\uff08\u5982nvidia-smi\uff09\uff0c\u5b58\u5728\u4e24\u5927\u7f3a\u9677\uff1a \u6570\u636e\u7c92\u5ea6\u7c97\u7cd9\uff1a\u4ec5\u80fd\u83b7\u53d6\u5168\u5c40\u7ea7\u522b\u7684\u5229\u7528\u7387\u7edf\u8ba1\uff0c\u65e0\u6cd5\u5173\u8054\u5230\u5177\u4f53\u8fdb\u7a0b\u6216\u5bb9\u5668\uff1b \u5ef6\u8fdf\u9ad8\uff1a\u91c7\u6837\u95f4\u9694\u901a\u5e38\u4e3a\u79d2\u7ea7\uff0c\u96be\u4ee5\u6355\u6349\u77ac\u65f6\u74f6\u9888\uff08\u5982\u5fae\u79d2\u7ea7\u6838\u51fd\u6570\u963b\u585e\uff09\u3002 2. eBPF\u7684\u89e3\u51b3\u65b9\u6848 eBPF\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u7a81\u7834\u4e0a\u8ff0\u9650\u5236\uff1a \u52a8\u6001\u63d2\u6869GPU\u9a71\u52a8\u51fd\u6570\uff1a\u5728\u5185\u6838\u6001\u62e6\u622aCUDA\u9a71\u52a8\u4e0e\u786c\u4ef6\u7684\u4ea4\u4e92\uff08\u5982cuMemAlloc\u3001cuLaunchKernel\uff09\uff0c\u8bb0\u5f55\u6bcf\u6b21\u8c03\u7528\u7684\u65f6\u95f4\u6233\u3001\u8fdb\u7a0bID\u53ca\u53c2\u6570\uff1b \u5173\u8054\u7528\u6237\u6001\u4e0e\u5185\u6838\u6001\u4e8b\u4ef6\uff1a\u7ed3\u5408uprobe\u8ffd\u8e2a\u5e94\u7528\u7a0b\u5e8f\u7684CUDA API\u8c03\u7528\uff08\u5982cudaMalloc\uff09\uff0c\u5f62\u6210\u4ece\u5e94\u7528\u4ee3\u7801\u5230\u786c\u4ef6\u6267\u884c\u7684\u5b8c\u6574\u8c03\u7528\u94fe\uff1b \u4f4e\u5f00\u9500\u6570\u636e\u805a\u5408\uff1a\u5728\u5185\u6838\u4e2d\u76f4\u63a5\u8fc7\u6ee4\u3001\u7edf\u8ba1\u4e8b\u4ef6\uff0c\u4ec5\u5c06\u7ed3\u679c\u4f20\u9012\u7ed9\u7528\u6237\u6001\uff0c\u907f\u514d\u6d77\u91cf\u539f\u59cb\u6570\u636e\u590d\u5236\u9020\u6210\u7684\u6027\u80fd\u635f\u8017\u3002 \u901a\u8fc7\u6302\u94a9NVIDIA\u9a71\u52a8\u4e2d\u7684\u5185\u5b58\u5206\u914d\u51fd\u6570\uff0ceBPF\u53ef\u5b9e\u65f6\u7edf\u8ba1\u6bcf\u4e2a\u5bb9\u5668\u7684\u663e\u5b58\u5360\u7528\uff0c\u5e76\u5728\u8d85\u9650\u65f6\u89e6\u53d1\u56de\u6536\u7b56\u7565\uff0c\u907f\u514d\u56e0\u5355\u5bb9\u5668\u8d44\u6e90\u6ee5\u7528\u5bfc\u81f4\u7684\u6574\u4f53\u6027\u80fd\u4e0b\u964d\u3002 \u4e09\u3001eBPF\u76d1\u63a7GPU\u7684\u6838\u5fc3\u6280\u672f\u8def\u5f84 1. \u5185\u6838\u6001\u63a2\u9488\u6ce8\u5165 eBPF\u7a0b\u5e8f\u901a\u8fc7kprobe\u6216tracepoint\u6302\u8f7d\u5230GPU\u9a71\u52a8\u7684\u5173\u952e\u51fd\u6570\u3002\u4ee5\u76d1\u63a7CUDA\u6838\u51fd\u6570\u6267\u884c\u4e3a\u4f8b\uff1a SEC(&quot;kprobe\/cuLaunchKernel&quot;) int trace_cu_launch(struct pt_regs *ctx) { u64 pid = bpf_get_current_pid_tgid(); u64 duration_ns = bpf_ktime_get_ns(); \/\/ \u8bb0\u5f55\u8fdb\u7a0bPID\u3001\u6838\u51fd\u6570\u540d\u79f0\u53ca\u542f\u52a8\u65f6\u95f4 bpf_map_update_elem(&amp;kernel_events, &amp;pid, &amp;duration_ns, BPF_ANY); return 0; } \u6b64\u7c7b\u7a0b\u5e8f\u53ef\u7cbe\u786e\u7edf\u8ba1\u6bcf\u4e2a\u6838\u51fd\u6570\u7684\u6267\u884c\u65f6\u957f\uff0c\u5e76\u5173\u8054\u5230\u53d1\u8d77\u8fdb\u7a0b\uff0c\u8fdb\u800c\u8bc6\u522b\u51fa\u4f4e\u6548\u7684GPU\u4ee3\u7801\u903b\u8f91\u3002 2. \u7528\u6237\u6001-\u5185\u6838\u6001\u6570\u636e\u534f\u540c eBPF Maps\u4f5c\u4e3a\u5185\u6838\u4e0e\u7528\u6237\u6001\u5171\u4eab\u7684\u6570\u636e\u7ed3\u6784\uff0c\u652f\u6301\u5b9e\u65f6\u4f20\u9012\u76d1\u63a7\u6307\u6807\u3002\u4f8b\u5982\uff0c\u901a\u8fc7Hash [&hellip;]<\/p>\n","protected":false},"author":4,"featured_media":33377,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[14],"tags":[],"class_list":["post-33376","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-product-news"],"_links":{"self":[{"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/posts\/33376","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\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/comments?post=33376"}],"version-history":[{"count":1,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/posts\/33376\/revisions"}],"predecessor-version":[{"id":33378,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/posts\/33376\/revisions\/33378"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/media\/33377"}],"wp:attachment":[{"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/media?parent=33376"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/categories?post=33376"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.yunshan.net\/blog\/wp-json\/wp\/v2\/tags?post=33376"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}