{"id":12700,"date":"2026-06-10T04:20:07","date_gmt":"2026-06-09T20:20:07","guid":{"rendered":"https:\/\/ab123.xyz\/?p=12700"},"modified":"2026-06-10T04:20:07","modified_gmt":"2026-06-09T20:20:07","slug":"kubernetes-gpu-%e8%b5%84%e6%ba%90%e8%b0%83%e5%ba%a6%e4%bc%98%e5%8c%96%e6%96%b9%e6%a1%88%ef%bc%9anvidia-gpu-operator-%e6%b7%b1%e5%ba%a6%e8%a7%a3%e6%9e%90","status":"publish","type":"post","link":"https:\/\/ab123.xyz\/?p=12700","title":{"rendered":"Kubernetes GPU \u8d44\u6e90\u8c03\u5ea6\u4f18\u5316\u65b9\u6848\uff1aNVIDIA GPU Operator \u6df1\u5ea6\u89e3\u6790"},"content":{"rendered":"<p>\u5728\u4e91\u539f\u751f\u65f6\u4ee3\uff0cKubernetes \u5df2\u6210\u4e3a\u5bb9\u5668\u7f16\u6392\u7684\u4e8b\u5b9e\u6807\u51c6\u3002\u7136\u800c\uff0c\u968f\u7740 AI \u8bad\u7ec3\u3001\u6df1\u5ea6\u5b66\u4e60\u63a8\u7406\u548c\u5927\u89c4\u6a21\u79d1\u5b66\u8ba1\u7b97\u7684\u666e\u53ca\uff0c\u5982\u4f55\u9ad8\u6548\u8c03\u5ea6\u548c\u7ba1\u7406 GPU \u8d44\u6e90\u6210\u4e3a\u8fd0\u7ef4\u56e2\u961f\u7684\u6838\u5fc3\u75db\u70b9\u3002NVIDIA GPU Operator \u4f5c\u4e3a\u5b98\u65b9\u63a8\u51fa\u7684\u667a\u80fd\u5de5\u5177\uff0c\u5f7b\u5e95\u6539\u53d8\u4e86 Kubernetes \u96c6\u7fa4\u4e2d GPU \u8d44\u6e90\u7684\u4f7f\u7528\u65b9\u5f0f\u3002\u8be5\u5de5\u5177\u901a\u8fc7\u81ea\u52a8\u5316\u9a71\u52a8\u5b89\u88c5\u3001\u8bbe\u5907\u63d2\u4ef6\u90e8\u7f72\u548c\u52a8\u6001\u8d44\u6e90\u5206\u914d\uff0c\u663e\u8457\u63d0\u5347\u4e86 GPU \u5229\u7528\u7387\u5e76\u964d\u4f4e\u4e86\u8fd0\u7ef4\u590d\u6742\u5ea6\u3002\u5b98\u65b9\u6700\u65b0\u7248\u672c\u5df2\u652f\u6301\u591a\u5b9e\u4f8b GPU\uff08MIG\uff09\u548c GPU \u65f6\u95f4\u5207\u7247\u529f\u80fd\uff0c\u4e3a\u6df7\u5408\u5de5\u4f5c\u8d1f\u8f7d\u63d0\u4f9b\u4e86\u5f39\u6027\u8c03\u5ea6\u80fd\u529b\u3002\u7acb\u5373\u8bbf\u95ee <a href=\"https:\/\/developer.nvidia.com\/gpu-operator\" target=\"_blank\">\u5b98\u65b9\u7f51\u7ad9<\/a> \u83b7\u53d6\u5b89\u88c5\u6307\u5357\u548c\u6700\u4f73\u5b9e\u8df5\u3002<\/p>\n<h2>\u6838\u5fc3\u529f\u80fd\u4e0e\u67b6\u6784\u4f18\u52bf<\/h2>\n<p>NVIDIA GPU Operator \u57fa\u4e8e Kubernetes Operator \u6a21\u5f0f\uff0c\u5c06 GPU \u76f8\u5173\u7684\u751f\u547d\u5468\u671f\u7ba1\u7406\u5c01\u88c5\u4e3a\u58f0\u660e\u5f0f API\u3002\u5176\u6838\u5fc3\u529f\u80fd\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u81ea\u52a8\u68c0\u6d4b\u5e76\u5b89\u88c5\u4e0e\u4e3b\u673a NVIDIA \u9a71\u52a8\u7248\u672c\u5339\u914d\u7684\u5bb9\u5668\u5316\u9a71\u52a8\u5305<\/li>\n<li>\u901a\u8fc7 Device Plugin \u6846\u67b6\u5411 kubelet \u6ce8\u518c GPU \u8bbe\u5907\uff0c\u5b9e\u73b0\u8282\u70b9\u7ea7\u8d44\u6e90\u4e0a\u62a5<\/li>\n<li>\u652f\u6301 GPU \u5206\u533a\u6280\u672f\uff08MIG\uff09\uff0c\u5141\u8bb8\u5c06\u4e00\u5757 A100 \u6216 H100 GPU \u5207\u5206\u4e3a\u591a\u4e2a\u72ec\u7acb\u7b97\u529b\u5355\u5143<\/li>\n<li>\u96c6\u6210 GPU \u6307\u6807\u66b4\u9732\u673a\u5236\uff0c\u4e0e Prometheus \u548c Grafana \u65e0\u7f1d\u5bf9\u63a5\uff0c\u5b9e\u73b0\u5b9e\u65f6\u76d1\u63a7<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u529f\u80fd\u4ece\u6839\u672c\u4e0a\u89e3\u51b3\u4e86\u4f20\u7edf\u624b\u52a8\u90e8\u7f72\u4e2d\u9a71\u52a8\u7248\u672c\u4e0d\u4e00\u81f4\u3001\u8d44\u6e90\u788e\u7247\u5316\u548c\u7ba1\u7406\u64cd\u4f5c\u5272\u88c2\u7684\u95ee\u9898\u3002<\/p>\n<h2>\u5173\u952e\u4f18\u52bf\uff1a\u8d44\u6e90\u5229\u7528\u7387\u4e0e\u5f39\u6027\u8c03\u5ea6<\/h2>\n<p>\u5728\u4f20\u7edf\u65b9\u6848\u4e2d\uff0cGPU \u8d44\u6e90\u901a\u5e38\u4ee5\u6574\u5361\u7c92\u5ea6\u5206\u914d\uff0c\u5bfc\u81f4\u5927\u91cf\u7b97\u529b\u95f2\u7f6e\u3002NVIDIA GPU Operator \u901a\u8fc7\u4ee5\u4e0b\u673a\u5236\u5b9e\u73b0\u7cbe\u7ec6\u5316\u7ba1\u7406\uff1a<\/p>\n<h3>MIG \u8d44\u6e90\u52a8\u6001\u5206\u914d<\/h3>\n<p>\u5bf9\u4e8e\u652f\u6301 MIG \u7684 GPU\uff0cOperator \u53ef\u4ee5\u81ea\u52a8\u521b\u5efa\u548c\u9500\u6bc1\u5206\u533a\uff0c\u5e76\u6839\u636e Pod \u7684\u8d44\u6e90\u8bf7\u6c42\uff08\u5982 1g.10gb \u89c4\u683c\uff09\u52a8\u6001\u7ed1\u5b9a\u7b97\u529b\u5207\u7247\u3002\u8fd9\u4f7f\u5f97\u591a\u79df\u6237\u73af\u5883\u4e0b\u7684\u5c0f\u578b\u63a8\u7406\u4efb\u52a1\u80fd\u5171\u4eab\u540c\u4e00\u7269\u7406 GPU\uff0c\u5c06\u5e73\u5747\u5229\u7528\u7387\u4ece 30% \u63d0\u5347\u81f3 70% \u4ee5\u4e0a\u3002<\/p>\n<h3>\u65f6\u95f4\u5207\u7247\u4e0e\u62a2\u5360\u5f0f\u8c03\u5ea6<\/h3>\n<p>\u9488\u5bf9\u4e0d\u652f\u6301 MIG \u7684\u65e7\u6b3e GPU\uff0cOperator \u5229\u7528 NVIDIA \u7684 Time-slicing \u63d2\u4ef6\uff0c\u5728\u591a\u4e2a Pod \u4e4b\u95f4\u6309\u65f6\u95f4\u7247\u8f6e\u8f6c\u5206\u914d GPU \u7b97\u529b\u3002\u7ed3\u5408 Kubernetes \u7684 PriorityClass \u4e0e Preemption\uff0c\u53ef\u786e\u4fdd\u9ad8\u4f18\u5148\u7ea7\u8bad\u7ec3\u4efb\u52a1\u4f18\u5148\u83b7\u5f97\u5b8c\u6574\u7b97\u529b\uff0c\u4f4e\u4f18\u5148\u7ea7\u4efb\u52a1\u5728\u7a7a\u95f2\u65f6\u6bb5\u8fd0\u884c\u3002<\/p>\n<h2>\u5e94\u7528\u573a\u666f\u4e0e\u5b9e\u8df5\u6307\u5357<\/h2>\n<p>\u8be5\u5de5\u5177\u5df2\u88ab\u5e7f\u6cdb\u5e94\u7528\u4e8e\u91d1\u878d\u98ce\u63a7\u6a21\u578b\u8bad\u7ec3\u3001\u81ea\u52a8\u9a7e\u9a76\u56fe\u50cf\u8bc6\u522b\u3001\u836f\u7269\u5206\u5b50\u6a21\u62df\u7b49\u9ad8\u6027\u80fd\u8ba1\u7b97\u9886\u57df\u3002\u5178\u578b\u90e8\u7f72\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\u5728 Kubernetes \u96c6\u7fa4\u4e2d\u5b89\u88c5 cert-manager \u4f5c\u4e3a\u8bc1\u4e66\u7ba1\u7406\u57fa\u7840\u7ec4\u4ef6<\/li>\n<li>\u901a\u8fc7 Helm Chart \u90e8\u7f72 GPU Operator\uff0c\u914d\u7f6e\u9a71\u52a8\u7248\u672c\u548c MIG \u7b56\u7565<\/li>\n<li>\u521b\u5efa\u542f\u7528 GPU \u7684\u5de5\u4f5c\u8d1f\u8f7d\uff08\u5982 TensorFlow \u8bad\u7ec3\u4efb\u52a1\uff09\uff0c\u5e76\u901a\u8fc7 resource.limits \u6307\u5b9a nvidia.com\/gpu \u6570\u91cf<\/li>\n<li>\u4f7f\u7528 kubectl top node \u7ed3\u5408 NVIDIA DCGM \u6307\u6807\u9a8c\u8bc1\u8d44\u6e90\u5206\u914d\u6548\u679c<\/li>\n<\/ul>\n<p>\u5b98\u65b9\u6587\u6863\u63d0\u4f9b\u4e86\u9488\u5bf9\u4e0d\u540c GPU \u578b\u53f7\u548c Kubernetes \u7248\u672c\u7684\u8be6\u7ec6\u6559\u7a0b\u3002\u5efa\u8bae\u751f\u4ea7\u73af\u5883\u914d\u5408 Kubernetes \u81ea\u5e26\u7684 Cluster Autoscaler \u4ee5\u53ca\u8282\u70b9\u6c60\u81ea\u52a8\u6269\u7f29\u5bb9\u7b56\u7565\uff0c\u5b9e\u73b0 GPU \u8d44\u6e90\u7684\u5f39\u6027\u4f38\u7f29\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u4e91\u539f\u751f\u65f6\u4ee3\uff0cKubernetes \u5df2\u6210\u4e3a\u5bb9\u5668\u7f16\u6392\u7684\u4e8b\u5b9e\u6807\u51c6\u3002\u7136\u800c\uff0c\u968f\u7740 AI \u8bad\u7ec3\u3001\u6df1\u5ea6\u5b66\u4e60\u63a8\u7406\u548c\u5927\u89c4\u6a21\u79d1\u5b66 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[4957,15447,15055,15445,15446],"class_list":["post-12700","post","type-post","status-publish","format-standard","hentry","category-4","tag-gpu","tag-k8s","tag-kubernetes-gpu","tag-nvidia-gpu-operator","tag-ai"],"_links":{"self":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/12700","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=12700"}],"version-history":[{"count":1,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/12700\/revisions"}],"predecessor-version":[{"id":12701,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/12700\/revisions\/12701"}],"wp:attachment":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=12700"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=12700"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=12700"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}