{"id":172,"date":"2026-09-03T00:06:15","date_gmt":"2026-09-02T16:06:15","guid":{"rendered":"https:\/\/www.xiaoxindong.cloud\/?p=172"},"modified":"2026-09-06T19:51:59","modified_gmt":"2026-09-06T11:51:59","slug":"gpu-%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2%ef%bc%88gpu-%e8%b5%84%e6%ba%90%e7%9b%91%e6%8e%a7%e4%b8%8e%e8%bf%90%e7%bb%b4%e8%af%8a%e6%96%ad%ef%bc%89","status":"publish","type":"post","link":"https:\/\/www.xiaoxindong.cloud\/?p=172","title":{"rendered":"GPU \u5927\u6a21\u578b\u672c\u5730\u90e8\u7f72\uff08GPU \u8d44\u6e90\u76d1\u63a7\u4e0e\u8fd0\u7ef4\u8bca\u65ad\uff09"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">\u672c\u7bc7\u805a\u7126 GPU \u786c\u4ef6\u5065\u5eb7\u8bca\u65ad\u3001\u8d44\u6e90\u76d1\u63a7\u4e0e\u53ef\u89c2\u6d4b\u4f53\u7cfb\u5efa\u8bbe\uff0c\u6838\u5fc3\u638c\u63e1<strong>DCGM \u5b98\u65b9\u8bca\u65ad\u5de5\u5177<\/strong>\u3001<strong>DCGM-Exporter \u76d1\u63a7\u539f\u7406<\/strong>\u3001<strong>K8s \u73af\u5883\u4e0b Prometheus+Grafana GPU \u76d1\u63a7\u6808\u90e8\u7f72<\/strong>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001NVIDIA DCGM \u5b98\u65b9\u8bca\u65ad\u5de5\u5177<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u5de5\u5177\u6982\u8ff0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">DCGM\uff08Data Center GPU Manager\uff09\u662f NVIDIA \u5b98\u65b9\u63a8\u51fa\u7684\u6570\u636e\u4e2d\u5fc3\u7ea7 GPU \u7ba1\u7406\u5de5\u5177\u96c6\uff0c\u63d0\u4f9b GPU \u884c\u4e3a\u76d1\u63a7\u3001\u914d\u7f6e\u7ba1\u7406\u3001\u5065\u5eb7\u8bca\u65ad\u3001\u7b56\u7565\u76d1\u7763\u80fd\u529b\uff0c\u662f GPU \u670d\u52a1\u5668\u8fd0\u7ef4\u3001\u6545\u969c\u6392\u67e5\u3001\u65b0\u673a\u9a8c\u6536\u7684\u6807\u51c6\u5de5\u5177\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u5b89\u88c5\u6b65\u9aa4<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code># 1. \u6dfb\u52a0CUDA\u5b98\u65b9\u5bc6\u94a5\nwget https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/ubuntu2204\/x86_64\/cuda-keyring_1.0-1_all.deb\ndpkg -i cuda-keyring_1.0-1_all.deb\n\n# 2. \u66f4\u65b0\u6e90\u5e76\u5b89\u88c5\napt update\napt install -y datacenter-gpu-manager\n\n# 3. \u542f\u52a8\u670d\u52a1\u5e76\u8bbe\u7f6e\u5f00\u673a\u81ea\u542f\nsystemctl start nvidia-dcgm\nsystemctl enable nvidia-dcgm\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u6838\u5fc3\u5b50\u547d\u4ee4\u4e0e\u5e38\u7528\u64cd\u4f5c<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u5b50\u547d\u4ee4<\/th><th>\u529f\u80fd<\/th><th>\u5e38\u7528\u64cd\u4f5c<\/th><\/tr><\/thead><tbody><tr><td><code>discovery<\/code><\/td><td>\u53d1\u73b0\u4e0e\u67e5\u8be2 GPU \u8bbe\u5907<\/td><td><code>dcgmi discovery -l<\/code> \u5217\u51fa\u6240\u6709 GPU<code>dcgmi discovery --gpuid 0 -i a<\/code> \u67e5\u770b\u5355\u5361\u8be6\u7ec6\u4fe1\u606f\uff08\u578b\u53f7\u3001PCI ID\u3001UUID\u3001\u5e8f\u5217\u53f7\u3001\u56fa\u4ef6\u7248\u672c\uff09<\/td><\/tr><tr><td><code>topo<\/code><\/td><td>\u67e5\u770b GPU \u62d3\u6251\u7ed3\u6784<\/td><td><code>dcgmi topo --gpuid 0<\/code> \u67e5\u770b GPU \u4e0e\u5176\u4ed6\u5361\u7684\u8fde\u63a5\u65b9\u5f0f\uff08PCIe \u4e3b\u673a\u6865\u3001PCIe \u4ea4\u6362\u673a\u3001CPU \u7ea7\u94fe\u8def\uff09<\/td><\/tr><tr><td><code>nvlink<\/code><\/td><td>\u68c0\u67e5 NVLink \u94fe\u8def\u72b6\u6001<\/td><td><code>dcgmi nvlink -s<\/code> \u67e5\u770b\u94fe\u8def\u72b6\u6001\u4e0e\u9519\u8bef\u8ba1\u6570\u72b6\u6001\u6807\u8bc6\uff1a<code>U<\/code>\u6b63\u5e38 \/ <code>D<\/code>\u672a\u5efa\u7acb \/ <code>X<\/code>\u7981\u7528 \/ <code>_<\/code>\u4e0d\u652f\u6301<\/td><\/tr><tr><td><code>health<\/code><\/td><td>GPU \u5065\u5eb7\u76d1\u63a7\u7ba1\u7406<\/td><td><code>dcgmi health -s a<\/code> \u5f00\u542f\u5168\u90e8\u5065\u5eb7\u76d1\u63a7\u5b50\u7cfb\u7edf<code>dcgmi health -c<\/code> \u67e5\u770b\u5f53\u524d\u5065\u5eb7\u72b6\u6001\u76d1\u63a7\u9879\uff1aPCIe\u3001NVLink\u3001\u5185\u5b58\u3001SM\u3001InfoROM\u3001\u6e29\u5ea6\u3001\u7535\u6e90\u3001\u9a71\u52a8<\/td><\/tr><tr><td><code>dmon<\/code><\/td><td>\u5b9e\u65f6\u6307\u6807\u76d1\u63a7<\/td><td><code>dcgmi dmon -e 319,320<\/code> \u68c0\u67e5 ECC \u9519\u8bef<code>dcgmi dmon -e 203,204<\/code> \u68c0\u67e5 GPU \u5229\u7528\u7387\u4e0e\u663e\u5b58\u5e26\u5bbd<\/td><\/tr><tr><td><code>diag<\/code><\/td><td>GPU \u8bca\u65ad\u4e0e\u538b\u6d4b\uff08\u9762\u8bd5\u9ad8\u9891\uff09<\/td><td>\u5206 4 \u4e2a\u8bca\u65ad\u7ea7\u522b\uff0c\u9010\u7ea7\u52a0\u6df1\u538b\u529b<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">4. DCGM \u8bca\u65ad\u7ea7\u522b<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7ea7\u522b<\/th><th>\u68c0\u67e5\u5185\u5bb9<\/th><th>\u8017\u65f6<\/th><th>\u538b\u529b<\/th><th>\u9002\u7528\u573a\u666f<\/th><\/tr><\/thead><tbody><tr><td>Level 1<\/td><td>GPU \u5b58\u5728\u6027\u3001\u9a71\u52a8\u72b6\u6001\u3001CUDA \u73af\u5883\u3001\u57fa\u7840\u901a\u4fe1\u68c0\u67e5<\/td><td>\u51e0\u79d2<\/td><td>\u65e0<\/td><td>\u65e5\u5e38\u5de1\u68c0\u3001\u73af\u5883\u9a8c\u8bc1<\/td><\/tr><tr><td>Level 2<\/td><td>Level1 \u5168\u90e8 + PCIe \u68c0\u67e5\u3001NVLink \u68c0\u67e5\u3001\u57fa\u7840\u8fd0\u7b97\u6d4b\u8bd5\u3001\u663e\u5b58\u68c0\u67e5<\/td><td>\u51e0\u5341\u79d2<\/td><td>\u8f7b\u5ea6<\/td><td>GPU \u6389\u5361\u6392\u67e5\u3001\u9a71\u52a8\u5f02\u5e38\u6392\u67e5\u3001\u8282\u70b9\u4ea4\u4ed8\u9a8c\u6536<\/td><\/tr><tr><td>Level 3<\/td><td>Level2 \u5168\u90e8 + SM \u538b\u529b\u6d4b\u8bd5\u3001\u663e\u5b58\u538b\u529b\u6d4b\u8bd5\u3001\u76ee\u6807\u529f\u8017\u6d4b\u8bd5\u3001ECC \u6df1\u5ea6\u68c0\u67e5\u3001\u8fd0\u7b97\u6b63\u786e\u6027\u9a8c\u8bc1<\/td><td>3~10 \u5206\u949f<\/td><td>\u9ad8\u8d1f\u8f7d<\/td><td>\u65b0\u673a\u4ea4\u4ed8\u9a8c\u6536\u3001\u786c\u4ef6\u6545\u969c\u6df1\u5ea6\u6392\u67e5\u3001\u8bad\u7ec3 \/ \u63a8\u7406\u5f02\u5e38\u5b9a\u4f4d<\/td><\/tr><tr><td>Level 4<\/td><td>\u66f4\u6df1\u5ea6\u7684\u786c\u4ef6\u7a33\u5b9a\u6027\u3001\u4f9b\u7535\u3001\u663e\u5b58\u52a3\u5316\u9a8c\u8bc1<\/td><td>\u66f4\u957f<\/td><td>\u6781\u9650<\/td><td>\u7591\u96be\u786c\u4ef6\u6545\u969c\u3001\u663e\u5b58\u7a33\u5b9a\u6027\u95ee\u9898\u4e13\u9879\u9a8c\u8bc1<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u751f\u4ea7\u73af\u5883\u6700\u5e38\u7528 Level 3 \u505a\u65b0\u673a\u538b\u6d4b\u4e0e\u6545\u969c\u6df1\u5ea6\u6392\u67e5\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001GPU \u5065\u5eb7\u5de1\u68c0 Shell \u811a\u672c<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u5b9a\u4f4d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u81ea\u52a8\u5316\u6279\u91cf\u6267\u884c GPU \u5065\u5eb7\u68c0\u67e5\u9879\uff0c\u8f93\u51fa\u5206\u7ea7\u544a\u8b66\u4e0e\u65e5\u5fd7\u62a5\u544a\uff0c\u9002\u5408\u65e5\u5e38\u6279\u91cf\u5de1\u68c0\u3001\u5b9a\u65f6\u5de1\u68c0\u573a\u666f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u6838\u5fc3\u8986\u76d6\u68c0\u67e5\u9879<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7cfb\u7edf\u57fa\u7840\u4fe1\u606f\uff1aOS \u7248\u672c\u3001\u5185\u6838\u3001CPU\u3001\u5185\u5b58<\/li>\n\n\n\n<li>NVIDIA \u9a71\u52a8\u68c0\u67e5\uff1a\u9a71\u52a8\u7248\u672c\u3001CUDA \u7248\u672c\u3001\u6a21\u5757\u52a0\u8f7d\u72b6\u6001<\/li>\n\n\n\n<li>GPU \u57fa\u672c\u4fe1\u606f\uff1a\u6570\u91cf\u3001\u578b\u53f7\u3001UUID\u3001PCI \u603b\u7ebf\u3001BIOS \u7248\u672c<\/li>\n\n\n\n<li>\u5065\u5eb7\u72b6\u6001\uff1a\u6e29\u5ea6\u3001\u529f\u8017\u3001\u98ce\u6247\u8f6c\u901f\u3001\u6027\u80fd\u72b6\u6001\uff08\u5206\u7ea7\u544a\u8b66\uff09<\/li>\n\n\n\n<li>\u6027\u80fd\u72b6\u6001\uff1aGPU \u5229\u7528\u7387\u3001\u663e\u5b58\u5229\u7528\u7387\u3001\u663e\u5b58\u4f7f\u7528\u91cf\u3001\u8ba1\u7b97\u6a21\u5f0f<\/li>\n\n\n\n<li>\u94fe\u8def\u72b6\u6001\uff1aPCIe \u901f\u7387 \/ \u5bbd\u5ea6\uff08\u662f\u5426\u964d\u901f\uff09\u3001NVLink \u6d3b\u52a8\u94fe\u8def\u6570<\/li>\n\n\n\n<li>\u8fdb\u7a0b\u5360\u7528\uff1a\u8fd0\u884c\u4e2d GPU \u8fdb\u7a0b\u7684 PID\u3001\u540d\u79f0\u3001\u663e\u5b58\u5360\u7528<\/li>\n\n\n\n<li>DCGM \u96c6\u6210\uff1a\u8c03\u7528 DCGM \u505a\u5feb\u901f\u8bca\u65ad<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u7279\u70b9<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u652f\u6301\u6e29\u5ea6\u3001\u5229\u7528\u7387\u9608\u503c\u81ea\u5b9a\u4e49\uff0c\u5206\u7ea7\u544a\u8b66\uff08\u8b66\u544a \/ \u4e34\u754c\uff09<\/li>\n\n\n\n<li>\u81ea\u52a8\u751f\u6210\u8be6\u7ec6\u65e5\u5fd7\u62a5\u544a\u4e0e\u6458\u8981\u6587\u4ef6<\/li>\n\n\n\n<li>\u9047\u5230\u5173\u952e\u9519\u8bef\u7acb\u5373\u9000\u51fa\uff0c\u9002\u5408\u81ea\u52a8\u5316\u6d41\u6c34\u7ebf\u8c03\u7528<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001K8s \u96c6\u7fa4 GPU \u76d1\u63a7\u4f53\u7cfb<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u6574\u4f53\u67b6\u6784<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u91c7\u7528\u6807\u51c6\u4e91\u539f\u751f\u53ef\u89c2\u6d4b\u6808\uff1a<strong>\u91c7\u96c6\u5c42 \u2192 \u5b58\u50a8\u5c42 \u2192 \u53ef\u89c6\u5316\u5c42 \u2192 \u544a\u8b66\u5c42<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u91c7\u96c6\u5c42<\/strong>\uff1aDCGM-Exporter\uff08GPU \u786c\u4ef6\u6307\u6807\uff09\u3001kube-state-metrics\uff08K8s \u5bf9\u8c61\u6307\u6807\uff09\u3001cAdvisor\uff08\u5bb9\u5668\u8d44\u6e90\u6307\u6807\uff09<\/li>\n\n\n\n<li><strong>\u5b58\u50a8\u5c42<\/strong>\uff1aPrometheus \u65f6\u5e8f\u6570\u636e\u5e93<\/li>\n\n\n\n<li><strong>\u53ef\u89c6\u5316\u5c42<\/strong>\uff1aGrafana \u4eea\u8868\u76d8<\/li>\n\n\n\n<li><strong>\u544a\u8b66\u5c42<\/strong>\uff1aAlertmanager + \u90ae\u4ef6 \/ \u9489\u9489 \/ \u4f01\u4e1a\u5fae\u4fe1\u901a\u77e5<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6570\u636e\u6d41\u5411\uff1aGPU \u786c\u4ef6 \u2192 DCGM \u670d\u52a1 \u2192 DCGM-Exporter \u2192 Prometheus \u91c7\u96c6\u5b58\u50a8 \u2192 Grafana \u53ef\u89c6\u5316 + \u544a\u8b66\u89c4\u5219\u89e6\u53d1<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. Prometheus \u90e8\u7f72\uff08K8s \u73af\u5883\uff09<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">\u90e8\u7f72\u6d41\u7a0b<\/h4>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u521b\u5efa<code>monitoring<\/code>\u4e13\u7528\u547d\u540d\u7a7a\u95f4<\/li>\n\n\n\n<li>\u7528 ConfigMap \u914d\u7f6e Prometheus \u6293\u53d6\u89c4\u5219\u4e0e\u5168\u5c40\u53c2\u6570<\/li>\n\n\n\n<li>\u521b\u5efa ClusterRole \u4e0e ServiceAccount\uff0c\u6388\u4e88\u96c6\u7fa4\u8d44\u6e90\u53d1\u73b0\u6743\u9650<\/li>\n\n\n\n<li>\u90e8\u7f72 Prometheus Deployment\uff0c\u6302\u8f7d\u914d\u7f6e\u4e0e\u6570\u636e\u5377<\/li>\n\n\n\n<li>\u901a\u8fc7 NodePort \u7c7b\u578b Service \u66b4\u9732\u670d\u52a1<\/li>\n<\/ol>\n\n\n\n<h4 class=\"wp-block-heading\">\u6838\u5fc3\u914d\u7f6e<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5168\u5c40\u6293\u53d6\u95f4\u9694\uff1a<code>scrape_interval: 15s<\/code><\/li>\n\n\n\n<li>\u652f\u6301 Kubernetes \u670d\u52a1\u53d1\u73b0\uff08<code>kubernetes_sd_configs<\/code>\uff09\uff0c\u81ea\u52a8\u53d1\u73b0 Pod\u3001Service\u3001Endpoint \u7b49\u76d1\u63a7\u76ee\u6807<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3. DCGM-Exporter(\u2b50)<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">\u5b9a\u4f4d<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u4e8e DCGM API \u5f00\u53d1\u7684 Prometheus \u6307\u6807\u5bfc\u51fa\u5668\uff0c\u5c06 NVIDIA GPU \u7684\u786c\u4ef6\u6307\u6807\u8f6c\u6362\u4e3a Prometheus \u6807\u51c6\u65f6\u5e8f\u683c\u5f0f\uff0c\u7531 NVIDIA \u5b98\u65b9\u7ef4\u62a4\uff1bK8s \u73af\u5883\u4e2d\u4ee5<strong>DaemonSet<\/strong>\u5f62\u5f0f\u90e8\u7f72\uff0c\u6bcf\u4e2a GPU \u8282\u70b9\u8fd0\u884c\u4e00\u4e2a\u5b9e\u4f8b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u6838\u5fc3\u7279\u70b9<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8986\u76d6 40 + \u9879 GPU \u6307\u6807\uff0c\u7ef4\u5ea6\u5168\u9762<\/li>\n\n\n\n<li>\u8d44\u6e90\u5f00\u9500\u8fdc\u4f4e\u4e8e\u5faa\u73af\u8c03\u7528 nvidia-smi<\/li>\n\n\n\n<li>\u5b98\u65b9\u7ef4\u62a4\uff0c\u4e0e\u65b0 GPU \u578b\u53f7\u517c\u5bb9\u6027\u5f3a<\/li>\n\n\n\n<li>\u539f\u751f\u652f\u6301\u5bb9\u5668\u5316\u4e0e K8s \u7f16\u6392\uff0c\u9002\u914d\u5927\u89c4\u6a21\u96c6\u7fa4<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u5de5\u4f5c\u539f\u7406<\/h4>\n\n\n\n<ol class=\"wp-block-list\">\n<li>DCGM \u901a\u8fc7 NVML \u63a5\u53e3\u4ece GPU \u9a71\u52a8\u91c7\u96c6\u786c\u4ef6\u6570\u636e<\/li>\n\n\n\n<li>DCGM-Exporter \u8c03\u7528 DCGM API \u8bfb\u53d6\u6307\u6807\uff0c\u8f6c\u6362\u4e3a Prometheus \u683c\u5f0f<\/li>\n\n\n\n<li>\u901a\u8fc7 9400 \u7aef\u53e3\u7684<code>\/metrics<\/code>\u63a5\u53e3\u66b4\u9732\u6307\u6807<\/li>\n\n\n\n<li>Prometheus \u901a\u8fc7\u670d\u52a1\u53d1\u73b0\u81ea\u52a8\u6293\u53d6\u6240\u6709\u8282\u70b9\u7684 GPU \u6307\u6807<\/li>\n<\/ol>\n\n\n\n<h4 class=\"wp-block-heading\">\u6838\u5fc3\u5e38\u7528\u6307\u6807<\/h4>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u6307\u6807\u540d\u79f0<\/th><th>\u63cf\u8ff0<\/th><th>\u5355\u4f4d<\/th><th>\u91cd\u8981\u6027<\/th><\/tr><\/thead><tbody><tr><td><code>DCGM_FI_DEV_GPU_UTIL<\/code><\/td><td>GPU \u6838\u5fc3\u5229\u7528\u7387<\/td><td>%<\/td><td>\u9ad8<\/td><\/tr><tr><td><code>DCGM_FI_DEV_FB_USED<\/code><\/td><td>\u5df2\u4f7f\u7528\u663e\u5b58<\/td><td>MiB<\/td><td>\u9ad8<\/td><\/tr><tr><td><code>DCGM_FI_DEV_FB_FREE<\/code><\/td><td>\u7a7a\u95f2\u663e\u5b58<\/td><td>MiB<\/td><td>\u9ad8<\/td><\/tr><tr><td><code>DCGM_FI_DEV_GPU_TEMP<\/code><\/td><td>GPU \u6838\u5fc3\u6e29\u5ea6<\/td><td>\u00b0C<\/td><td>\u9ad8<\/td><\/tr><tr><td><code>DCGM_FI_DEV_POWER_USAGE<\/code><\/td><td>\u5f53\u524d\u5b9e\u9645\u529f\u8017<\/td><td>W<\/td><td>\u9ad8<\/td><\/tr><tr><td><code>DCGM_FI_DEV_XID_ERRORS<\/code><\/td><td>XID \u9519\u8bef\u8ba1\u6570<\/td><td>\u8ba1\u6570<\/td><td>\u9ad8<\/td><\/tr><tr><td><code>DCGM_FI_DEV_UNCORRECTABLE_ERRORS<\/code><\/td><td>\u4e0d\u53ef\u7ea0\u6b63 ECC \u9519\u8bef<\/td><td>\u8ba1\u6570<\/td><td>\u9ad8<\/td><\/tr><tr><td><code>DCGM_FI_DEV_CORRECTABLE_ERRORS<\/code><\/td><td>\u53ef\u7ea0\u6b63 ECC \u9519\u8bef<\/td><td>\u8ba1\u6570<\/td><td>\u4e2d<\/td><\/tr><tr><td><code>DCGM_FI_DEV_NVLINK_BANDWIDTH_TOTAL<\/code><\/td><td>NVLink \u603b\u5e26\u5bbd<\/td><td>KB\/s<\/td><td>\u9ad8<\/td><\/tr><tr><td><code>DCGM_FI_DEV_SM_CLOCK<\/code><\/td><td>SM \u6d41\u5904\u7406\u5668\u65f6\u949f<\/td><td>MHz<\/td><td>\u4e2d<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">\u90e8\u7f72\u4e0e\u96c6\u6210<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code># \u90e8\u7f72DaemonSet\nkubectl apply -f dcgm-exporter.yaml\n\n# Prometheus\u914d\u7f6e\u65b0\u589e\u6293\u53d6\u4efb\u52a1\n- job_name: 'dcgm-exporter'\n  kubernetes_sd_configs:\n  - role: endpoints\n    namespaces:\n      names: &#91;default]\n  relabel_configs:\n  - source_labels: &#91;__meta_kubernetes_service_name]\n    action: keep\n    regex: dcgm-exporter\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4. Grafana \u53ef\u89c6\u5316<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">\u90e8\u7f72<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5 Deployment+NodePort \u65b9\u5f0f\u90e8\u7f72\uff0c\u5bb9\u5668\u7aef\u53e3 3000 \u6620\u5c04\u5230\u8282\u70b9\u7aef\u53e3 32000\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u914d\u7f6e\u6b65\u9aa4<\/h4>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u8fdb\u5165 Grafana \u2192 \u8fde\u63a5 \u2192 \u6570\u636e\u6e90 \u2192 \u6dfb\u52a0 Prometheus \u6570\u636e\u6e90<\/li>\n\n\n\n<li>\u586b\u5165 Prometheus \u670d\u52a1\u5730\u5740\uff0c\u4fdd\u5b58\u5e76\u6d4b\u8bd5\u8fde\u901a\u6027<\/li>\n\n\n\n<li>\u4eea\u8868\u677f \u2192 \u5bfc\u5165 \u2192 \u8f93\u5165\u5b98\u65b9\u4eea\u8868\u76d8 ID <code>12239<\/code><\/li>\n\n\n\n<li>\u9009\u62e9 Prometheus \u6570\u636e\u6e90\uff0c\u70b9\u51fb\u5bfc\u5165\uff0c\u4e00\u952e\u751f\u6210\u5b8c\u6574 GPU \u76d1\u63a7\u9762\u677f<\/li>\n<\/ol>\n\n\n\n<h4 class=\"wp-block-heading\">\u6838\u5fc3\u76d1\u63a7\u9762\u677f<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">GPU \u6e29\u5ea6\u3001GPU \u5e73\u5747\u6e29\u5ea6\u3001\u529f\u8017\u3001SM \u65f6\u949f\u9891\u7387\u3001GPU \u5229\u7528\u7387\u3001\u5f20\u91cf\u6838\u5fc3\u4f7f\u7528\u7387\u3001\u663e\u5b58\u4f7f\u7528\u91cf\u3001ECC \u9519\u8bef\u8ba1\u6570\u3001NVLink \u5e26\u5bbd\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5. \u544a\u8b66\u89c4\u5219\u914d\u7f6e<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5178\u578b GPU \u544a\u8b66\u89c4\u5219\u793a\u4f8b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u544a\u8b66\u540d\u79f0<\/th><th>\u89e6\u53d1\u6761\u4ef6<\/th><th>\u6301\u7eed\u65f6\u95f4<\/th><th>\u4e25\u91cd\u7ea7\u522b<\/th><th>\u8bf4\u660e<\/th><\/tr><\/thead><tbody><tr><td>GPUHighUtilization<\/td><td><code>DCGM_FI_DEV_GPU_UTIL &gt; 95<\/code><\/td><td>1m<\/td><td>warning<\/td><td>\u5229\u7528\u7387\u6301\u7eed\u8fc7\u9ad8\uff0c\u53ef\u80fd\u5b58\u5728\u7b97\u529b\u74f6\u9888<\/td><\/tr><tr><td>GPUHighTemperature<\/td><td><code>DCGM_FI_DEV_GPU_TEMP &gt; 75<\/code><\/td><td>1m<\/td><td>warning<\/td><td>\u6e29\u5ea6\u8fc7\u9ad8\uff0c\u5b58\u5728\u964d\u9891\u98ce\u9669<\/td><\/tr><tr><td>GPUMemoryNearlyFull<\/td><td><code>DCGM_FI_DEV_FB_FREE \/ DCGM_FI_DEV_FB_TOTAL * 100 &lt; 5<\/code><\/td><td>1m<\/td><td>warning<\/td><td>\u663e\u5b58\u4e0d\u8db3\uff0c\u53ef\u80fd\u51fa\u73b0 OOM<\/td><\/tr><tr><td>GPUXidErrors<\/td><td><code>delta(DCGM_FI_DEV_XID_ERRORS[5m]) &gt; 0<\/code><\/td><td>\u5373\u65f6<\/td><td>critical<\/td><td>\u9a71\u52a8\u5d29\u6e83\u6216 GPU \u786c\u4ef6\u94fe\u8def\u5f02\u5e38<\/td><\/tr><tr><td>GPUECCErrors<\/td><td><code>delta(DCGM_FI_DEV_UNCORRECTABLE_ERRORS[1h]) &gt; 0<\/code><\/td><td>\u5373\u65f6<\/td><td>critical<\/td><td>\u663e\u5b58\u786c\u4ef6\u6545\u969c\uff0c\u9700\u7acb\u5373\u5904\u7406<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u6c47\u603b<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>GPU \u670d\u52a1\u5668\u4e0a\u7ebf\u524d\u7684\u5b8c\u6574\u5065\u5eb7\u68c0\u67e5\u6d41\u7a0b\u662f\u4ec0\u4e48\uff1f<\/strong> \u7b54\uff1a\u5148\u505a\u57fa\u7840\u68c0\u67e5\uff1a\u7528<code>nvidia-smi<\/code>\u786e\u8ba4 GPU \u6570\u91cf\u3001\u6e29\u5ea6\u3001\u529f\u8017\u3001ECC \u72b6\u6001\uff1b\u7528<code>nvidia-smi topo -m<\/code>\u68c0\u67e5\u62d3\u6251\uff1b\u7528<code>nvidia-smi nvlink -s<\/code>\u68c0\u67e5 NVLink \u94fe\u8def\u3002\u518d\u505a\u6df1\u5ea6\u9a8c\u8bc1\uff1a\u7528<code>dcgmi discovery<\/code>\u786e\u8ba4\u8bbe\u5907\u4fe1\u606f\u5b8c\u6574\uff1b\u5f00\u542f\u5168\u91cf\u5065\u5eb7\u76d1\u63a7\uff1b\u8fd0\u884c<code>dcgmi diag -r 3<\/code>\u9ad8\u7ea7\u538b\u529b\u8bca\u65ad\uff0c\u786e\u4fdd\u65e0\u786c\u4ef6\u6545\u969c\u3001\u65e0\u4e0d\u53ef\u7ea0\u6b63 ECC\u3001PCIe\/NVLink \u65e0\u964d\u901f\u3002<\/li>\n\n\n\n<li><strong>DCGM \u8bca\u65ad\u5206\u51e0\u4e2a\u7ea7\u522b\uff0c\u5206\u522b\u9002\u7528\u4ec0\u4e48\u573a\u666f\uff1f<\/strong> \u7b54\uff1a\u5171 4 \u4e2a\u7ea7\u522b\u3002Level 1 \u65e0\u538b\u529b\uff0c\u7528\u4e8e\u65e5\u5e38\u5de1\u68c0\uff1bLevel 2 \u8f7b\u5ea6\u538b\u529b\uff0c\u7528\u4e8e\u6389\u5361\u6392\u67e5\u3001\u8282\u70b9\u9a8c\u6536\uff1bLevel 3 \u6ee1\u8d1f\u8f7d\u538b\u529b\uff0c\u7528\u4e8e\u65b0\u673a\u4ea4\u4ed8\u3001\u786c\u4ef6\u6545\u969c\u6df1\u5ea6\u6392\u67e5\uff1bLevel 4 \u6781\u9650\u6df1\u5ea6\uff0c\u7528\u4e8e\u7591\u96be\u663e\u5b58\u3001\u4f9b\u7535\u95ee\u9898\u4e13\u9879\u9a8c\u8bc1\u3002<\/li>\n\n\n\n<li><strong>DCGM-Exporter \u7684\u5de5\u4f5c\u539f\u7406\u662f\u4ec0\u4e48\uff1f<\/strong> \u7b54\uff1aDCGM-Exporter \u57fa\u4e8e DCGM \u670d\u52a1\uff0c\u901a\u8fc7 NVML \u63a5\u53e3\u91c7\u96c6 GPU \u786c\u4ef6\u6307\u6807\uff0c\u8f6c\u6362\u4e3a Prometheus \u6807\u51c6\u65f6\u5e8f\u683c\u5f0f\uff0c\u901a\u8fc7 9400 \u7aef\u53e3\u66b4\u9732<code>\/metrics<\/code>\u63a5\u53e3\u3002K8s \u4e2d\u4ee5 DaemonSet \u90e8\u7f72\u5728\u6bcf\u4e2a GPU \u8282\u70b9\uff0cPrometheus \u901a\u8fc7\u670d\u52a1\u53d1\u73b0\u81ea\u52a8\u6293\u53d6\u6240\u6709\u8282\u70b9\u6307\u6807\uff0c\u5b9e\u73b0\u96c6\u7fa4\u7ea7\u7edf\u4e00\u76d1\u63a7\u3002<\/li>\n\n\n\n<li><strong>XID \u9519\u8bef\u548c ECC \u9519\u8bef\u5206\u522b\u662f\u4ec0\u4e48\uff0c\u600e\u4e48\u5904\u7406\uff1f<\/strong> \u7b54\uff1a\n<ul class=\"wp-block-list\">\n<li>XID \u9519\u8bef\uff1a\u9a71\u52a8\u5c42\u9762\u9519\u8bef\uff0c\u901a\u5e38\u7531\u9a71\u52a8\u5d29\u6e83\u3001GPU \u6389\u5361\u3001PCIe \u94fe\u8def\u5f02\u5e38\u5bfc\u81f4\uff1b\u51fa\u73b0\u589e\u91cf XID \u65f6\uff0c\u4f18\u5148\u6392\u67e5\u9a71\u52a8\u7248\u672c\u3001Fabric Manager\u3001PCIe \u94fe\u8def\u72b6\u6001\u3002<\/li>\n\n\n\n<li>ECC \u9519\u8bef\uff1a\u663e\u5b58\u6821\u9a8c\u9519\u8bef\uff0c\u5206\u53ef\u7ea0\u6b63\uff08CE\uff09\u548c\u4e0d\u53ef\u7ea0\u6b63\uff08UE\uff09\uff1b\u5c11\u91cf CE \u53ef\u8bb0\u5f55\u57fa\u7ebf\u89c2\u5bdf\uff0c\u6301\u7eed\u589e\u957f\u9700\u6392\u67e5\u663e\u5b58\u52a3\u5316\uff1bUE \u4e3a\u4e25\u91cd\u786c\u4ef6\u9519\u8bef\uff0c\u9700\u7acb\u5373\u8fc1\u79fb\u4e1a\u52a1\uff0c\u505a\u786c\u4ef6\u8bca\u65ad\uff0c\u8003\u8651 RMA \u8fd4\u4fee\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>GPU \u76d1\u63a7\u7684\u56db\u5927\u7c7b\u6838\u5fc3\u6307\u6807\u662f\u4ec0\u4e48\uff1f<\/strong> \u7b54\uff1a\u2460\u8fd0\u884c\u6027\u80fd\uff1aGPU \u5229\u7528\u7387\u3001\u5f20\u91cf\u6838\u5fc3\u5229\u7528\u7387\u3001SM \u65f6\u949f\uff1b\u2461\u5b58\u50a8\u72b6\u6001\uff1a\u663e\u5b58\u4f7f\u7528\u7387\u3001\u5df2\u7528 \/ \u7a7a\u95f2\u91cf\uff1b\u2462\u5065\u5eb7\u72b6\u6001\uff1a\u6e29\u5ea6\u3001\u529f\u8017\u3001\u98ce\u6247\u8f6c\u901f\uff1b\u2463\u9519\u8bef\u8ba1\u6570\uff1aXID \u9519\u8bef\u3001ECC \u9519\u8bef\u3001NVLink \u9519\u8bef\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u7bc7\u805a\u7126 GPU \u786c\u4ef6\u5065\u5eb7\u8bca\u65ad\u3001\u8d44\u6e90\u76d1\u63a7\u4e0e\u53ef\u89c2\u6d4b\u4f53\u7cfb\u5efa\u8bbe\uff0c\u6838\u5fc3\u638c\u63e1DCGM \u5b98\u65b9\u8bca\u65ad\u5de5\u5177\u3001DCGM-Expor [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-172","post","type-post","status-publish","format-standard","hentry","category-6"],"_links":{"self":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/172","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=172"}],"version-history":[{"count":2,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/172\/revisions"}],"predecessor-version":[{"id":184,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/172\/revisions\/184"}],"wp:attachment":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=172"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=172"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=172"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}