{"id":1622,"date":"2020-06-21T18:05:43","date_gmt":"2020-06-21T10:05:43","guid":{"rendered":"http:\/\/39.97.231.10\/?p=1622"},"modified":"2020-06-21T18:05:43","modified_gmt":"2020-06-21T10:05:43","slug":"kubernetes-%e8%b0%83%e5%ba%a6gups","status":"publish","type":"post","link":"http:\/\/www.weijing.co\/?p=1622","title":{"rendered":"kubernetes \u8c03\u5ea6 GPUs"},"content":{"rendered":"<h1>\u8c03\u5ea6 GPUs<\/h1>\n<p>Kubernetes \u652f\u6301\u5bf9\u8282\u70b9\u4e0a\u7684 AMD \u548c NVIDA GPU \u8fdb\u884c\u7ba1\u7406\uff0c\u76ee\u524d\u5904\u4e8e<strong>\u5b9e\u9a8c<\/strong>\u72b6\u6001\u3002\u5bf9 NVIDIA GPU \u7684\u652f\u6301\u5728 v1.6 \u4e2d\u52a0\u5165\uff0c\u5df2\u7ecf\u7ecf\u5386\u4e86\u591a\u6b21\u4e0d\u5411\u540e\u517c\u5bb9\u7684\u8fed\u4ee3\u3002\u800c\u5bf9 AMD GPU \u7684\u652f\u6301\u5219\u5728 v1.9 \u4e2d\u901a\u8fc7\u00a0\u8bbe\u5907\u63d2\u4ef6\u00a0\u52a0\u5165\u3002<br \/>\n\u8fd9\u4e2a\u9875\u9762\u4ecb\u7ecd\u4e86\u7528\u6237\u5982\u4f55\u5728\u4e0d\u540c\u7684 Kubernetes \u7248\u672c\u4e2d\u4f7f\u7528 GPU\uff0c\u4ee5\u53ca\u5f53\u524d\u5b58\u5728\u7684\u4e00\u4e9b\u9650\u5236\u3002<\/p>\n<h2 id=\\\"\u4ece-v1-8-\u8d77\\\">\u4ece v1.8 \u8d77<\/h2>\n<p><strong>\u4ece 1.8 \u7248\u672c\u5f00\u59cb\uff0c\u6211\u4eec\u63a8\u8350\u901a\u8fc7\u00a0\u8bbe\u5907\u63d2\u4ef6\u00a0\u7684\u65b9\u5f0f\u6765\u4f7f\u7528 GPU\u3002<\/strong><br \/>\n\u5728 1.10 \u7248\u672c\u4e4b\u524d\uff0c\u4e3a\u4e86\u901a\u8fc7\u8bbe\u5907\u63d2\u4ef6\u5f00\u542f GPU \u7684\u652f\u6301\uff0c\u6211\u4eec\u9700\u8981\u5728\u7cfb\u7edf\u4e2d\u5c06\u00a0<code>DevicePlugins<\/code>\u00a0\u8fd9\u4e00\u7279\u6027\u5f00\u5173\u663e\u5f0f\u5730\u8bbe\u7f6e\u4e3a true\uff1a<code>--feature-gates=\\\"DevicePlugins=true\\\"<\/code>\u3002\u4e0d\u8fc7\uff0c \u4ece 1.10 \u7248\u672c\u5f00\u59cb\uff0c\u6211\u4eec\u5c31\u4e0d\u9700\u8981\u8fd9\u4e00\u6b65\u9aa4\u4e86\u3002<br \/>\n\u63a5\u7740\u4f60\u9700\u8981\u5728\u4e3b\u673a\u8282\u70b9\u4e0a\u5b89\u88c5\u5bf9\u5e94\u5382\u5546\u7684 GPU \u9a71\u52a8\u5e76\u8fd0\u884c\u5bf9\u5e94\u5382\u5546\u7684\u8bbe\u5907\u63d2\u4ef6 (AMD\u3001NVIDIA)\u3002<br \/>\n\u5f53\u4e0a\u9762\u7684\u6761\u4ef6\u90fd\u6ee1\u8db3\uff0cKubernetes \u5c06\u4f1a\u66b4\u9732\u00a0<code>nvidia.com\/gpu<\/code>\u00a0\u6216\u00a0<code>amd.com\/gpu<\/code>\u00a0\u6765\u4f5c\u4e3a \u4e00\u79cd\u53ef\u8c03\u5ea6\u7684\u8d44\u6e90\u3002<br \/>\n\u4f60\u4e5f\u80fd\u901a\u8fc7\u50cf\u8bf7\u6c42\u00a0<code>cpu<\/code>\u00a0\u6216\u00a0<code>memory<\/code>\u00a0\u4e00\u6837\u8bf7\u6c42\u00a0<code>&lt;vendor&gt;.com\/gpu<\/code>\u00a0\u6765\u5728\u5bb9\u5668\u4e2d\u4f7f\u7528 GPU\u3002\u7136\u800c\uff0c\u5f53\u4f60\u8981\u901a\u8fc7\u6307\u5b9a\u8d44\u6e90\u8bf7\u6c42\u6765\u4f7f\u7528 GPU \u65f6\uff0c\u5b58\u5728\u7740\u4ee5\u4e0b\u51e0\u70b9\u9650\u5236\uff1a<\/p>\n<ul>\n<li>GPU \u4ec5\u4ec5\u652f\u6301\u5728\u00a0<code>limits<\/code>\u00a0\u90e8\u5206\u88ab\u6307\u5b9a\uff0c\u8fd9\u8868\u660e\uff1a\n<ul>\n<li>\u4f60\u53ef\u4ee5\u4ec5\u4ec5\u6307\u5b9a GPU \u7684\u00a0<code>limits<\/code>\u00a0\u5b57\u6bb5\u800c\u4e0d\u5fc5\u987b\u6307\u5b9a\u00a0<code>requests<\/code>\u00a0\u5b57\u6bb5\uff0c\u56e0\u4e3a Kubernetes \u4f1a\u9ed8\u8ba4\u4f7f\u7528 limit \u5b57\u6bb5\u7684\u503c\u6765\u4f5c\u4e3a request \u5b57\u6bb5\u7684\u9ed8\u8ba4\u503c\u3002<\/li>\n<li>\u4f60\u80fd\u540c\u65f6\u6307\u5b9a GPU \u7684\u00a0<code>limits<\/code>\u00a0\u548c\u00a0<code>requests<\/code>\u00a0\u5b57\u6bb5\uff0c\u4f46\u8fd9\u4e24\u4e2a\u503c\u5fc5\u987b\u76f8\u7b49\u3002<\/li>\n<li>\u4f60\u4e0d\u80fd\u4ec5\u4ec5\u6307\u5b9a GPU \u7684\u00a0<code>request<\/code>\u00a0\u5b57\u6bb5\u800c\u4e0d\u6307\u5b9a\u00a0<code>limits<\/code>\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u5bb9\u5668\uff08\u4ee5\u53ca pod\uff09\u5e76\u4e0d\u4f1a\u5171\u4eab GPU\uff0c\u4e5f\u4e0d\u5b58\u5728\u5bf9 GPU \u7684\u8fc7\u91cf\u4f7f\u7528\u3002<\/li>\n<li>\u6bcf\u4e00\u4e2a\u5bb9\u5668\u80fd\u591f\u8bf7\u6c42\u4e00\u4e2a\u6216\u591a\u4e2a GPU\u3002\u7136\u800c\u53ea\u8bf7\u6c42\u4e00\u4e2a GPU \u7684\u4e00\u90e8\u5206\u662f\u4e0d\u5141\u8bb8\u7684\u3002<\/li>\n<\/ul>\n<p>\u4e0b\u9762\u662f\u4e00\u4e2a\u4f8b\u5b50:<\/p>\n<div>\n<pre>apiVersion: v1\nkind: Pod\nmetadata:\n  name: cuda-vector-add\nspec:\n  restartPolicy: OnFailure\n  containers:\n    - name: cuda-vector-add\n      # https:\/\/github.com\/kubernetes\/kubernetes\/blob\/v1.7.11\/test\/images\/nvidia-cuda\/Dockerfile\n      image: \\\"k8s.gcr.io\/cuda-vector-add:v0.1\\\"\n      resources:\n        limits:\n          nvidia.com\/gpu: 1 # requesting 1 GPU\n<\/pre>\n<\/div>\n<h3 id=\\\"\u90e8\u7f72-amd-gpu-\u8bbe\u5907\u63d2\u4ef6\\\">\u90e8\u7f72 AMD GPU \u8bbe\u5907\u63d2\u4ef6<\/h3>\n<p>\u5b98\u65b9\u7684 AMD GPU \u8bbe\u5907\u63d2\u4ef6\u00a0\u6709\u4ee5\u4e0b\u8981\u6c42\uff1a<\/p>\n<ul>\n<li>Kubernetes \u8282\u70b9\u5fc5\u987b\u9884\u5148\u5b89\u88c5 AMD GPU \u7684 Linux \u9a71\u52a8\u3002<\/li>\n<\/ul>\n<p>\u5982\u679c\u4f60\u7684\u96c6\u7fa4\u5df2\u7ecf\u542f\u52a8\u5e76\u4e14\u6ee1\u8db3\u4e0a\u8ff0\u8981\u6c42\u7684\u8bdd\uff0c\u53ef\u4ee5\u8fd9\u6837\u90e8\u7f72 AMD \u8bbe\u5907\u63d2\u4ef6\uff1a<\/p>\n<pre># \u9488\u5bf9 Kubernetes v1.9\nkubectl create -f https:\/\/raw.githubusercontent.com\/RadeonOpenCompute\/k8s-device-plugin\/r1.9\/k8s-ds-amdgpu-dp.yaml\n# \u9488\u5bf9 Kubernetes v1.10\nkubectl create -f https:\/\/raw.githubusercontent.com\/RadeonOpenCompute\/k8s-device-plugin\/r1.10\/k8s-ds-amdgpu-dp.yaml\n<\/pre>\n<p>\u8bf7\u5230\u00a0RadeonOpenCompute\/k8s-device-plugin\u00a0\u62a5\u544a\u6709\u5173\u6b64\u8bbe\u5907\u63d2\u4ef6\u7684\u95ee\u9898\u3002<\/p>\n<h3 id=\\\"\u90e8\u7f72-nvidia-gpu-\u8bbe\u5907\u63d2\u4ef6\\\">\u90e8\u7f72 NVIDIA GPU \u8bbe\u5907\u63d2\u4ef6<\/h3>\n<p>\u5bf9\u4e8e NVIDIA GPUs\uff0c\u76ee\u524d\u5b58\u5728\u4e24\u79cd\u8bbe\u5907\u63d2\u4ef6\u7684\u5b9e\u73b0\uff1a<\/p>\n<h4 id=\\\"\u5b98\u65b9\u7684-nvidia-gpu-\u8bbe\u5907\u63d2\u4ef6\\\">\u5b98\u65b9\u7684 NVIDIA GPU \u8bbe\u5907\u63d2\u4ef6<\/h4>\n<p><a href=\\\"https:\/\/github.com\/NVIDIA\/k8s-device-plugin\\\">\u5b98\u65b9\u7684 NVIDIA GPU \u8bbe\u5907\u63d2\u4ef6<\/a>\u00a0\u6709\u4ee5\u4e0b\u8981\u6c42:<\/p>\n<ul>\n<li>Kubernetes \u7684\u8282\u70b9\u5fc5\u987b\u9884\u5148\u5b89\u88c5\u4e86 NVIDIA \u9a71\u52a8<\/li>\n<li>Kubernetes \u7684\u8282\u70b9\u5fc5\u987b\u9884\u5148\u5b89\u88c5\u00a0<a href=\\\"https:\/\/github.com\/NVIDIA\/nvidia-docker\\\">nvidia-docker 2.0<\/a><\/li>\n<li>Docker \u7684<a href=\\\"https:\/\/github.com\/NVIDIA\/k8s-device-plugin#preparing-your-gpu-nodes\\\">\u9ed8\u8ba4\u8fd0\u884c\u65f6<\/a>\u5fc5\u987b\u8bbe\u7f6e\u4e3a nvidia-container-runtime\uff0c\u800c\u4e0d\u662f runc<\/li>\n<li>NVIDIA \u9a71\u52a8\u7248\u672c ~= 361.93<\/li>\n<\/ul>\n<p>\u5982\u679c\u4f60\u7684\u96c6\u7fa4\u5df2\u7ecf\u542f\u52a8\u5e76\u4e14\u6ee1\u8db3\u4e0a\u8ff0\u8981\u6c42\u7684\u8bdd\uff0c\u53ef\u4ee5\u8fd9\u6837\u90e8\u7f72 NVIDIA \u8bbe\u5907\u63d2\u4ef6\uff1a<\/p>\n<pre><code># \u9488\u5bf9 Kubernetes v1.8\nkubectl create -f https:\/\/raw.githubusercontent.com\/NVIDIA\/k8s-device-plugin\/v1.8\/nvidia-device-plugin.yml\n# \u9488\u5bf9 Kubernetes v1.9\nkubectl create -f https:\/\/raw.githubusercontent.com\/NVIDIA\/k8s-device-plugin\/v1.9\/nvidia-device-plugin.yml\n<\/code><\/pre>\n<h4 id=\\\"gce-\u4e2d\u4f7f\u7528\u7684-nvidia-gpu-\u8bbe\u5907\u63d2\u4ef6\\\">GCE \u4e2d\u4f7f\u7528\u7684 NVIDIA GPU \u8bbe\u5907\u63d2\u4ef6<\/h4>\n<p>GCE \u4f7f\u7528\u7684 NVIDIA GPU \u8bbe\u5907\u63d2\u4ef6\u00a0\u5e76\u4e0d\u8981\u6c42\u4f7f\u7528 nvidia-docker\uff0c\u5e76\u4e14\u5bf9\u4e8e\u4efb\u4f55\u5b9e\u73b0\u4e86 Kubernetes CRI \u7684\u5bb9\u5668\u8fd0\u884c\u65f6\uff0c\u90fd\u5e94\u8be5\u80fd\u591f\u4f7f\u7528\u3002\u8fd9\u4e00\u5b9e\u73b0\u5df2\u7ecf\u5728\u00a0Container-Optimized OS\u00a0\u4e0a\u8fdb\u884c\u4e86\u6d4b\u8bd5\uff0c\u5e76\u4e14\u5728 1.9 \u7248\u672c\u4e4b\u540e\u4f1a\u6709\u5bf9\u4e8e Ubuntu \u7684\u5b9e\u9a8c\u6027\u4ee3\u7801\u3002<br \/>\n\u5728\u4f60 1.12 \u7248\u672c\u7684\u96c6\u7fa4\u4e0a\uff0c\u4f60\u80fd\u4f7f\u7528\u4e0b\u9762\u7684\u547d\u4ee4\u6765\u5b89\u88c5 NVIDIA \u9a71\u52a8\u4ee5\u53ca\u8bbe\u5907\u63d2\u4ef6\uff1a<\/p>\n<pre># \u5728\u5bb9\u5668\u4f18\u5316\u7684\u64cd\u4f5c\u7cfb\u7edf\u4e0a\u5b89\u88c5 NVIDIA \u9a71\u52a8:\nkubectl create -f https:\/\/raw.githubusercontent.com\/GoogleCloudPlatform\/container-engine-accelerators\/stable\/daemonset.yaml\n# \u5728 Ubuntu \u4e0a\u5b89\u88c5 NVIDIA \u9a71\u52a8 (\u5b9e\u9a8c\u6027\u8d28):\nkubectl create -f https:\/\/raw.githubusercontent.com\/GoogleCloudPlatform\/container-engine-accelerators\/stable\/nvidia-driver-installer\/ubuntu\/daemonset.yaml\n# \u5b89\u88c5\u8bbe\u5907\u63d2\u4ef6:\nkubectl create -f https:\/\/raw.githubusercontent.com\/kubernetes\/kubernetes\/release-1.12\/cluster\/addons\/device-plugins\/nvidia-gpu\/daemonset.yaml\n<\/pre>\n<h2 id=\\\"\u96c6\u7fa4\u5185\u5b58\u5728\u4e0d\u540c\u7c7b\u578b\u7684-nvidia-gpu\\\">\u96c6\u7fa4\u5185\u5b58\u5728\u4e0d\u540c\u7c7b\u578b\u7684 NVIDIA GPU<\/h2>\n<p>\u5982\u679c\u96c6\u7fa4\u5185\u90e8\u7684\u4e0d\u540c\u8282\u70b9\u4e0a\u6709\u4e0d\u540c\u7c7b\u578b\u7684 NVIDIA GPU\uff0c\u90a3\u4e48\u4f60\u53ef\u4ee5\u4f7f\u7528\u00a0\u8282\u70b9\u6807\u7b7e\u548c\u8282\u70b9\u9009\u62e9\u5668\u00a0\u6765\u5c06 pod \u8c03\u5ea6\u5230\u5408\u9002\u7684\u8282\u70b9\u4e0a\u3002<br \/>\n\u4f8b\u5982\uff1a<\/p>\n<div>\n<pre># \u4e3a\u4f60\u7684\u8282\u70b9\u52a0\u4e0a\u5b83\u4eec\u6240\u62e5\u6709\u7684\u52a0\u901f\u5668\u7c7b\u578b\u7684\u6807\u7b7e\nkubectl label nodes &lt;node-with-k80&gt; accelerator=nvidia-tesla-k80\nkubectl label nodes &lt;node-with-p100&gt; accelerator=nvidia-tesla-p100\n<\/pre>\n<\/div>\n<p>\u5bf9\u4e8e AMD GPUs\uff0c\u60a8\u53ef\u4ee5\u90e8\u7f72\u00a0\u8282\u70b9\u6807\u7b7e\u5668\uff0c\u5b83\u4f1a\u81ea\u52a8\u7ed9\u8282\u70b9\u6253\u4e0a GPU \u5c5e\u6027\u6807\u7b7e\u3002\u76ee\u524d\u652f\u6301\u7684\u5c5e\u6027\uff1a<\/p>\n<ul>\n<li>\u8bbe\u5907 ID (-device-id)<\/li>\n<li>VRAM \u5927\u5c0f (-vram)<\/li>\n<li>SIMD \u6570\u91cf(-simd-count)<\/li>\n<li>\u8ba1\u7b97\u5355\u4f4d\u6570\u91cf(-cu-count)<\/li>\n<li>\u56fa\u4ef6\u548c\u7279\u6027\u7248\u672c (-firmware)<\/li>\n<li>GPU \u7cfb\u5217\uff0c\u4e24\u4e2a\u5b57\u6bcd\u7684\u9996\u5b57\u6bcd\u7f29\u5199(-family)\n<ul>\n<li>SI &#8211; Southern Islands<\/li>\n<li>CI &#8211; Sea Islands<\/li>\n<li>KV &#8211; Kaveri<\/li>\n<li>VI &#8211; Volcanic Islands<\/li>\n<li>CZ &#8211; Carrizo<\/li>\n<li>AI &#8211; Arctic Islands<\/li>\n<li>RV &#8211; Raven<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u793a\u4f8b:<\/p>\n<pre>$ kubectl describe node cluster-node-23\nName:               cluster-node-23\nRoles:              &lt;none&gt;\nLabels:             beta.amd.com\/gpu.cu-count.64=1\n                    beta.amd.com\/gpu.device-id.6860=1\n                    beta.amd.com\/gpu.family.AI=1\n                    beta.amd.com\/gpu.simd-count.256=1\n                    beta.amd.com\/gpu.vram.16G=1\n                    beta.kubernetes.io\/arch=amd64\n                    beta.kubernetes.io\/os=linux\n                    kubernetes.io\/hostname=cluster-node-23\nAnnotations:        kubeadm.alpha.kubernetes.io\/cri-socket: \/var\/run\/dockershim.sock\n                    node.alpha.kubernetes.io\/ttl: 0\n......\n<\/pre>\n<p>\u5728 pod \u7684 spec \u5b57\u6bb5\u4e2d\u6307\u5b9a GPU \u7684\u7c7b\u578b\uff1a<\/p>\n<div>\n<pre>apiVersion: v1\nkind: Pod\nmetadata:\n  name: cuda-vector-add\nspec:\n  restartPolicy: OnFailure\n  containers:\n    - name: cuda-vector-add\n      # https:\/\/github.com\/kubernetes\/kubernetes\/blob\/v1.7.11\/test\/images\/nvidia-cuda\/Dockerfile\n      image: \\\"k8s.gcr.io\/cuda-vector-add:v0.1\\\"\n      resources:\n        limits:\n          nvidia.com\/gpu: 1\n  nodeSelector:\n    accelerator: nvidia-tesla-p100 # or nvidia-tesla-k80 etc.\n<\/pre>\n<\/div>\n<p>\u8fd9\u80fd\u591f\u4fdd\u8bc1 pod \u80fd\u591f\u88ab\u8c03\u5ea6\u5230\u4f60\u6240\u6307\u5b9a\u7c7b\u578b\u7684 GPU \u7684\u8282\u70b9\u4e0a\u53bb\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8c03\u5ea6 GPUs Kubernetes \u652f\u6301\u5bf9\u8282\u70b9\u4e0a\u7684 AMD \u548c NVIDA GPU \u8fdb\u884c\u7ba1\u7406\uff0c\u76ee\u524d\u5904\u4e8e\u5b9e\u9a8c\u72b6 &hellip;<\/p>\n<p class=\"read-more\"> <a class=\"\" href=\"http:\/\/www.weijing.co\/?p=1622\"> <span class=\"screen-reader-text\">kubernetes \u8c03\u5ea6 GPUs<\/span> \u67e5\u770b\u5168\u6587 &raquo;<\/a><\/p>\n","protected":false},"author":4,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"default","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"default","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":""},"categories":[104,123],"tags":[],"uagb_featured_image_src":{"full":false,"thumbnail":false,"medium":false,"medium_large":false,"large":false,"1536x1536":false,"2048x2048":false,"mailpoet_newsletter_max":false,"woocommerce_thumbnail":false,"woocommerce_single":false,"woocommerce_gallery_thumbnail":false,"shop_catalog":false,"shop_single":false,"shop_thumbnail":false},"uagb_author_info":{"display_name":"weijing","author_link":"http:\/\/www.weijing.co\/?author=4"},"uagb_comment_info":0,"uagb_excerpt":"\u8c03\u5ea6 GPUs Kubernetes \u652f\u6301\u5bf9\u8282\u70b9\u4e0a\u7684 AMD \u548c NVIDA GPU \u8fdb\u884c\u7ba1\u7406\uff0c\u76ee\u524d\u5904\u4e8e\u5b9e\u9a8c\u72b6&hellip;","_links":{"self":[{"href":"http:\/\/www.weijing.co\/index.php?rest_route=\/wp\/v2\/posts\/1622"}],"collection":[{"href":"http:\/\/www.weijing.co\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.weijing.co\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.weijing.co\/index.php?rest_route=\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"http:\/\/www.weijing.co\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1622"}],"version-history":[{"count":0,"href":"http:\/\/www.weijing.co\/index.php?rest_route=\/wp\/v2\/posts\/1622\/revisions"}],"wp:attachment":[{"href":"http:\/\/www.weijing.co\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1622"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.weijing.co\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1622"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.weijing.co\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1622"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}