{"id":271,"date":"2021-03-11T15:34:06","date_gmt":"2021-03-11T07:34:06","guid":{"rendered":"\/?p=271"},"modified":"2021-03-11T15:34:06","modified_gmt":"2021-03-11T07:34:06","slug":"41-kubernetes-gpu%e7%ae%a1%e7%90%86%e5%92%8cdevice-plugin%e6%9c%ba%e5%88%b6","status":"publish","type":"post","link":"http:\/\/xinblog.ltd\/?p=271","title":{"rendered":"41.Kubernetes GPU\u7ba1\u7406\u548cDevice Plugin\u673a\u5236"},"content":{"rendered":"<p>\u4f34\u968f\u7740AlphaGO\u7684\u8d70\u7ea2\u548cTensorFlow\u9879\u76ee\u7684\u5f02\u519b\u7a81\u8d77,\u4e00\u573a\u540d\u4e3aAI\u7684\u6280\u672f\u9769\u547d\u4ece\u5b66\u672f\u754c\u8513\u5ef6\u5230\u4e86\u5de5\u4e1a\u754c<\/p>\n<p>\u5f53\u7136,\u673a\u5668\u5b66\u4e60\u6216\u8005\u4eba\u5de5\u667a\u80fd,\u5e76\u4e0d\u662f\u5565\u65b0\u9c9c\u73a9\u610f,\u4e0d\u8fc7\u5e26\u7ed9Kubernetes\u7684,\u5c31\u662f\u5728Kubernetes\u96c6\u7fa4\u4e0a\u8fd0\u884cTensorFlow\u7b49\u673a\u5668\u5b66\u4e60\u6846\u67b6\u9700\u8981\u7684\u8bad\u7ec3\u548c\u670d\u52a1\u4efb\u52a1,\u9664\u4e86\u524d\u9762\u8bf4\u7684Job,Operator\u7b49\u79bb\u7ebf\u4f5c\u4e1a\u7ba1\u7406\u9700\u8981\u7528\u5230\u7684\u7f16\u6392\u6982\u5ff5\u4e4b\u5916,\u8fd8\u6709\u5c31\u662f\u5bf9GPU\u7b49\u786c\u4ef6\u52a0\u901f\u8bbe\u5907\u7ba1\u7406\u7684\u652f\u6301<\/p>\n<p>\u5bf9\u4e8e\u4e91\u4e0a\u7684\u7528\u6237,\u5728GPU\u7684\u652f\u6301\u4e0a,\u57fa\u672c\u7684\u8bc9\u6c42\u5f88\u7b80\u5355,\u53ea\u8981\u5728Pod\u7684YAML\u4e2d,\u58f0\u660e\u67d0\u4e2a\u5bb9\u5668\u9700\u8981\u7684GPU\u4e2a\u6570,\u90a3\u4e48Kubernetes\u521b\u5efa\u7684\u5bb9\u5668\u5c31\u5e94\u8be5\u51fa\u73b0\u5bf9\u5e94\u7684GPU\u8bbe\u5907,\u4ee5\u53ca\u5bf9\u5e94\u7684\u9a71\u52a8\u76ee\u5f55<\/p>\n<p>\u4ee5NVDIA\u7684GPU\u8bbe\u5907\u4e3a\u4f8b,\u5bb9\u5668\u88ab\u521b\u5efa\u540e,\u8fd9\u4e2a\u5bb9\u5668\u4e2d\u4f1a\u51fa\u73b0GPU\u8bbe\u5907\u6982\u5ff5\u548cGPU\u7684\u9a71\u52a8\u76ee\u5f55<\/p>\n<p>\u4e0d\u8fc7,Kubernetes\u5e76\u6ca1\u6709\u4e3aGPU\u5728Pod\u7684API\u5bf9\u8c61\u8bbe\u7f6e\u4e00\u4e2a\u8d44\u6e90\u7c7b\u578b\u5b57\u6bb5,\u800c\u662f\u4f7f\u7528\u4e86\u4e00\u4e2a\u540d\u4e3aExtended Resource ER\u7684\u7279\u6b8a\u5b57\u6bb5\u6765\u8d1f\u8d23\u4f20\u9012GPU\u7684\u4fe1\u606f,\u5982\u4e0b<\/p>\n<table>\n<tbody>\n<tr>\n<td>apiVersion: v1<\/p>\n<p>kind: Pod<\/p>\n<p>metadata:<\/p>\n<p>name: cuda-vector-add<\/p>\n<p>spec:<\/p>\n<p>restartPolicy: OnFailure<\/p>\n<p>containers:<\/p>\n<p>&#8211; name: cuda-vector-add<\/p>\n<p>image: &#8220;k8s.gcr.io\/cuda-vector-add:v0.1&#8221;<\/p>\n<p>resources:<\/p>\n<p>limits:<\/p>\n<p>nvidia.com\/gpu: 1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6211\u4eec\u5728resources\u5b57\u6bb5\u4e2d,\u5b9a\u4e49\u4e86\u5b57\u6bb5\u540d\u4e3anvidia.com\/gpu,\u503c\u662f1,\u8fd9\u4e2aPod\u58f0\u660e\u8981\u4f7f\u7528\u4e00\u4e2aNVDIA\u7c7b\u578b\u7684GPU<\/p>\n<p>\u800c\u5728kube-scheduler\u4e2d,\u5e76\u4e0d\u5173\u5fc3\u8fd9\u4e2a\u5b57\u6bb5\u7684\u5177\u4f53\u542b\u4e49,\u800c\u662f\u5728\u8ba1\u7b97\u7684\u65f6\u5019,\u76f4\u63a5\u51cf\u53bb\u6216\u8005\u52a0\u4e0a\u5bf9\u5e94\u7684\u6570\u503c,Extended Resource,\u5c31\u662fKubernetes\u4e3a\u7528\u6237\u8bbe\u7f6e\u7684\u4e00\u79cd\u5bf9\u81ea\u5b9a\u4e49\u8d44\u6e90\u7684\u652f\u6301<\/p>\n<p>\u5728Kube-scheduler\u91cc\u9762,\u5e76\u4e0d\u5173\u5fc3\u8fd9\u4e2a\u5b57\u6bb5\u5177\u4f53\u542b\u4e49,\u53ea\u4f1a\u5728\u8ba1\u7b97\u7684\u65f6\u5019,\u4e00\u5f8b\u5c06\u8c03\u5ea6\u5668\u91cc\u4fdd\u5b58\u7684\u7c7b\u578b\u7684\u53ef\u7528\u91cf,\u76f4\u63a5\u51cf\u53bbPod\u58f0\u660e\u7684\u6570\u503c<\/p>\n<p>\u5f53\u7136,\u4e3a\u4e86\u53ef\u4ee5\u8ba1\u7b97,\u5bbf\u4e3b\u673a\u8282\u70b9\u672c\u8eab,\u9700\u8981\u80fd\u591f\u5411API Server\u6c47\u62a5\u8fd9\u4e2a\u7c7b\u578b\u8d44\u6e90\u7684\u53ef\u7528\u6570\u91cf,\u5728Kubernetes\u4e2d,\u5404\u4e2a\u7c7b\u578b\u7684\u8d44\u6e90\u53ef\u7528\u91cf,\u5c31\u662fNode\u4e2dStatus\u5b57\u6bb5\u53ef\u7528\u5185\u5bb9,<\/p>\n<table>\n<tbody>\n<tr>\n<td>apiVersion: v1<\/p>\n<p>kind: Node<\/p>\n<p>metadata:<\/p>\n<p>name: node-1<\/p>\n<p>&#8230;<\/p>\n<p>Status:<\/p>\n<p>Capacity:<\/p>\n<p>cpu:\u00a0 2<\/p>\n<p>memory:\u00a0 2049008Ki<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6211\u4eec\u5b9a\u4e49\u4e86Status\u4e2d\u7684cpu\u6570\u91cf\u4e3a2\u4e2a<\/p>\n<p>\u6211\u4eec\u8fd8\u9700\u8981\u5b9a\u4e49\u5bf9\u5e94\u7684Node\u8282\u70b9\u4e2d\u7684GPU\u6570\u636e,\u6211\u4eec\u9700\u8981\u4f7f\u7528PATCH API\u6765\u5bf9\u8fd9\u4e2aNode\u5bf9\u8c61\u8fdb\u884c\u66f4\u65b0\u64cd\u4f5c<\/p>\n<p>PATCH\u64cd\u4f5c\u5b8c\u6210\u540e,\u5c31\u53ef\u4ee5\u770b\u5230Node\u7684Status\u53d8\u4e3a\u4e86\u5982\u4e0b<\/p>\n<table>\n<tbody>\n<tr>\n<td>apiVersion: v1<\/p>\n<p>kind: Node<\/p>\n<p>&#8230;<\/p>\n<p>Status:<\/p>\n<p>Capacity:<\/p>\n<p>cpu:\u00a0 2<\/p>\n<p>memory:\u00a0 2049008Ki<\/p>\n<p>nvidia.com\/gpu: 1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8fd9\u6837,\u7f13\u5b58\u4e2d\u4e5f\u4f1a\u66f4\u65b0\u4e86\u7f13\u5b58\u4e2d\u8bb0\u5f55\u7684GPU\u7c7b\u578b\u7684\u6570\u91cf\u4e3a1<\/p>\n<p>\u5728Kubernetes\u7684GPU\u652f\u6301\u65b9\u6848\u4e2d,\u6211\u4eec\u4e0d\u9700\u8981\u505a\u5173\u4e8eExtended Resource\u7684\u914d\u7f6e\u64cd\u4f5c,\u5bf9\u6240\u6709\u786c\u4ef6\u52a0\u901f\u7ba1\u7406\u7684\u529f\u80fd,\u90fd\u662f\u4ea4\u7ed9\u4e86\u53eb\u505aDevice Plugin\u7684\u63d2\u4ef6\u8fdb\u884c\u8d1f\u8d23\u7684,\u8fd9\u5176\u4e2d,\u5c31\u5305\u62ec\u4e86\u5bf9\u8fd9\u4e2a\u786c\u4ef6\u7684Extended Resource\u8fdb\u884c\u6c47\u62a5\u7684\u903b\u8f91<\/p>\n<p>Kubernetes\u7684Device Plugin\u673a\u5236,\u53ef\u4ee5\u663e\u793a\u4e3a\u5982\u4e0b\u6240\u793a\u7684\u793a\u610f\u56fe<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" width=\"1640\" height=\"808\" class=\"wp-image-272\" src=\"\/wp-content\/uploads\/2021\/03\/unnamed-file-95.png\" alt=\"\u56fe\u7247\" srcset=\"http:\/\/xinblog.ltd\/wp-content\/uploads\/2021\/03\/unnamed-file-95.png 1640w, http:\/\/xinblog.ltd\/wp-content\/uploads\/2021\/03\/unnamed-file-95-300x148.png 300w, http:\/\/xinblog.ltd\/wp-content\/uploads\/2021\/03\/unnamed-file-95-1024x505.png 1024w, http:\/\/xinblog.ltd\/wp-content\/uploads\/2021\/03\/unnamed-file-95-768x378.png 768w, http:\/\/xinblog.ltd\/wp-content\/uploads\/2021\/03\/unnamed-file-95-1536x757.png 1536w\" sizes=\"(max-width: 1640px) 100vw, 1640px\" \/><\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u786c\u4ef6\u8bbe\u5907,\u90fd\u9700\u8981\u548c\u5bf9\u5e94\u7684Device Plugin\u8fdb\u884c\u7ba1\u7406,\u8fd9\u4e9bDevice Plugin,\u90fd\u4f1a\u901a\u8fc7GRPC\u7684\u65b9\u5f0f,\u548ckubelet\u8fde\u63a5\u8d77\u6765,\u4ee5NVDIA GPU\u4e3a\u4f8b,\u5bf9\u5e94\u7684\u63d2\u4ef6\u53eb\u505a,NVDIA GPU DEVICE plugin<\/p>\n<p>\u8fd9\u4e2aDevice plugin\u4f1a\u901a\u8fc7\u4e00\u4e2aListAndWatch\u7684API,\u5b9a\u671f\u5411kubelet\u6c47\u62a5\u8fd9\u4e2aNode\u4e0a\u7684GPU\u5217\u8868,\u5728\u4f8b\u5b50\u4e2d,\u4e00\u5171\u6709\u4e09\u4e2aGPU,\u8fd9\u6837,Kubelet\u4e2d\u5c31\u4f1a\u83b7\u53d6\u5230\u8fd9\u4f1a\u6709\u4e09\u4e2aGPU\u5217\u8868,\u4f46\u662fListAndWatch\u4e0a\u62a5\u7684\u4fe1\u606f,\u53ea\u6709\u672c\u673a\u7684GPU ID list,\u5e76\u6ca1\u6709GPU\u8bbe\u5907\u672c\u8eab\u7684\u4fe1\u606f.<\/p>\n<p>\u6545\u5f53\u4e00\u4e2aPod\u60f3\u8981\u4f7f\u7528GPU\u7684\u65f6\u5019,\u53ea\u9700\u8981\u5728\u672c\u6587\u4e00\u5f00\u59cb\u7ed9\u51fa\u7684\u4f8b\u5b50\u4e00\u6837,\u5728Pod\u7684limits\u5b57\u6bb5\u58f0\u660e nvidia.com\/gpu: 1,\u90a3\u4e48\u63a5\u4e0b\u6765,Kubernetes\u5c31\u4f1a\u5728\u7f13\u5b58\u4e2d,\u5bfb\u627eGPU\u6570\u91cf\u6ee1\u8db3\u6761\u4ef6\u7684Node,\u5c06\u7f13\u5b58\u7684GPU\u6570\u91cf\u51cf\u4e00\u540e\u8fdb\u884c\u7ed1\u5b9a.<\/p>\n<p>\u8c03\u5ea6\u6210\u529f\u540e\u7684Pod\u4fe1\u606f,\u5c31\u4f1a\u88abkubelet\u62ff\u6765\u505a\u5bb9\u5668\u64cd\u4f5c,\u5f53kubelet\u53d1\u73b0Pod\u9700\u8981\u4e00\u4e2aGPU,\u5c31\u4f1a\u5411Dervice Plugin\u53d1\u8d77\u4e00\u4e2aAllocate()\u8bf7\u6c42,\u8bf7\u6c42\u7684\u53c2\u6570,\u6b63\u662f\u5206\u914d\u7ed9\u5bb9\u5668\u7684\u8bbe\u5907ID\u5217\u8868<\/p>\n<p>\u7136\u540eDevice Plugin\u6536\u5230\u4e86Allocate\u8bf7\u6c42\u540e,\u5c31\u4f1a\u6839\u636ekubelet\u4f20\u6765\u7684\u8bbe\u5907ID,\u4eceDevice Plugin\u627e\u5230\u8fd9\u4e9b\u8bbe\u5907\u5bf9\u5e94\u7684\u8bbe\u5907\u8def\u5f84\u548c\u9a71\u52a8\u76ee\u5f55,\u5728NVIDIA Device Plugin\u7684\u5b9e\u73b0\u91cc,\u5b9a\u671f\u8bbf\u95eenvida docke\u63d2\u4ef6,\u83b7\u53d6\u672c\u673a\u4e0a\u7684GPU\u4fe1\u606f<\/p>\n<p>\u800c\u88abGPU\u7684\u8bbe\u5907\u76ee\u5f55\u548c\u9a71\u52a8\u76ee\u5f55\u4fe1\u606f\u90fd\u4f1a\u63a8\u7ed9kubelet,kubelet\u5c31\u5b8c\u6210\u4e86\u4e00\u4e2a\u5bb9\u5668\u5206\u914dGPU\u7684\u64cd\u4f5c,\u63a5\u4e0b\u6765,kubelet\u5c31\u4f1a\u5c06\u8fd9\u4e9b\u4fe1\u606f\u7ed9\u4e0e\u521b\u5efa\u5bb9\u5668\u7684CRI\u8bf7\u6c42,\u53d1\u7ed9Docker,\u8ba9Docker\u5b8c\u6210\u6302\u8f7d<\/p>\n<p>\u8fd9\u6837\u5c31\u662f\u4e3a\u4e00\u4e2aPod\u5206\u914d\u4e00\u4e2aGPU\u7684\u6d41\u7a0b<\/p>\n<p>\u5bf9\u4e8e\u5176\u4ed6\u7c7b\u578b\u786c\u4ef6\u6765\u8bf4,\u60f3\u8981\u5728Kubernetes\u6240\u7ba1\u7406\u7684\u5bb9\u5668\u4f7f\u7528\u8fd9\u4e9b\u786c\u4ef6\u7684\u8bdd,\u9700\u8981\u9075\u5faa\u4e0a\u8ff0Device Plugin\u7684\u6d41\u7a0b\u6765\u5b9e\u73b0\u5982\u4e0b\u7684Allocate\u548cListAndWatch API:<\/p>\n<table>\n<tbody>\n<tr>\n<td>\u00a0 service DevicePlugin {<\/p>\n<p>\/\/ ListAndWatch returns a stream of List of Devices<\/p>\n<p>\/\/ Whenever a Device state change or a Device disappears, ListAndWatch<\/p>\n<p>\/\/ returns the new list<\/p>\n<p>rpc ListAndWatch(Empty) returns (stream ListAndWatchResponse) {}<\/p>\n<p>\/\/ Allocate is called during container creation so that the Device<\/p>\n<p>\/\/ Plugin can run device specific operations and instruct Kubelet<\/p>\n<p>\/\/ of the steps to make the Device available in the container<\/p>\n<p>rpc Allocate(AllocateRequest) returns (AllocateResponse) {}<\/p>\n<p>}<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Kubernetes\u5b9e\u73b0\u4e86\u5f88\u591a\u7684\u786c\u4ef6\u63d2\u4ef6,\u6bd4\u5982FPGA\u7b49<\/p>\n<p>\u6211\u4eec\u5728\u672c\u7247\u4e2d\u8bf4\u4e86Kubernetes\u5bf9\u4e8eGPU\u7684\u7ba1\u7406\u65b9\u5f0f,\u4ee5\u53ca\u4f7f\u7528\u7684Device Plugin\u673a\u5236,Device Plugin\u7684\u8bbe\u8ba1,\u957f\u671f\u662f\u4ee5Google Cloud\u7528\u6237\u9700\u6c42\u4e3a\u4e3b\u5bfc\u7684,\u6240\u4ee5\u8fd8\u662f\u6709\u4e00\u5b9a\u533a\u522b\u7684<\/p>\n<p>GPU\u7b49\u786c\u4ef6\u8bbe\u5907\u7684\u8c03\u5ea6,\u4ea4\u7ed9\u4e86kubelet,\u4f46\u662fkubelet\u53ea\u662f\u4ece\u81ea\u5df1\u7684Node\u4e2d,\u6839\u636e\u8bbe\u5907\u6570\u76ee\u53bb\u5b9a\u4e49\u4f7f\u7528,\u5982\u679cGPU\u7684\u4f7f\u7528\u8c03\u5ea6\u4e0a\u8981\u6c42\u6bd4\u8f83\u590d\u6742,\u5982\u679c\u8bbe\u5907\u662f\u5f02\u6784\u7684,\u4e0d\u80fd\u7b80\u5355\u7684\u4f7f\u7528\u6570\u76ee\u53bb\u5b9a\u4e49\u7684\u65f6\u5019,\u5982\u679c\u60f3\u8981\u4f7f\u7528\u6307\u5b9a\u7684GPU\u7684\u8bdd,\u7b80\u5355\u7684Device Plugin\u5c31\u4e0d\u80fd\u5904\u7406\u4e86<\/p>\n<p>\u5728\u5f88\u591a\u573a\u666f\u4e0b,\u6211\u4eec\u5e0c\u671b\u5728\u8c03\u5ea6\u5668\u8fdb\u884c\u8c03\u5ea6\u7684\u65f6\u5019,\u53ef\u4ee5\u6839\u636e\u6574\u4e2a\u96c6\u7fa4\u4e2d\u7684\u67d0\u79cd\u786c\u4ef6\u8bbe\u5907\u7684\u5168\u5c40\u5206\u5e03,\u505a\u51fa\u6700\u4f73\u7684\u8c03\u5ea6<\/p>\n<p>\u6b64\u5916,\u4e0a\u8ff0Device Plugin\u7684\u8bbe\u8ba1,\u4e5f\u4f7f\u5f97Kubernetes\u4e2d,\u7f3a\u5c11\u4e00\u79cd\u80fd\u591f\u5bf9Device\u8fdb\u884c\u63cf\u8ff0\u7684API\u5bf9\u8c61,\u5982\u679c\u786c\u4ef6\u5bf9\u8c61\u5f88\u590d\u6742\u5e76\u4e14Pod\u5f88\u5173\u5fc3,\u90a3\u4e48Device Plugin\u662f\u6ca1\u6709\u529e\u6cd5\u652f\u6301\u7684<\/p>\n<p>Device Plugin\u7684\u8bbe\u8ba1\u548c\u5b9e\u73b0\u4e2d,Google \u7684\u5de5\u7a0b\u5e08\u4e5f\u4e0d\u613f\u610f\u6dfb\u52a0\u4e00\u4e9b\u6269\u5c55\u6027\u7684\u53c2\u6570,\u5bfc\u81f4\u65e0\u6cd5\u5f88\u597d\u5730\u81ea\u5b9a\u4e49<\/p>\n<p>\u90a3\u4e48\u5982\u4f55\u5bf9\u8fd9\u4e9bDevice Plugin\u8fdb\u884c\u6539\u52a8\u5462?<\/p>\n<p>\u73b0\u5728\u7684\u8bbe\u8ba1\u662f\u5426\u5df2\u7ecf\u591f\u7528\u4e86\u5417<\/p>\n<p>\u73b0\u5728\u7684\u8d44\u6e90\u5fc5\u7136\u4e0d\u591f\u7528,\u56e0\u4e3a\u53ea\u80fd\u6309\u7167\u6574\u6570\u7c7b\u578b\u7684\u5206\u914d,\u5bfc\u81f4\u5f88\u591a\u65f6\u5019,\u4e0d\u80fd\u5171\u4eab\u8d44\u6e90<\/p>\n<p>\u65e0\u6cd5\u505a\u5230\u6309\u9700\u4fee\u6539api<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f34\u968f\u7740AlphaGO\u7684\u8d70\u7ea2\u548cT [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"_links":{"self":[{"href":"http:\/\/xinblog.ltd\/index.php?rest_route=\/wp\/v2\/posts\/271"}],"collection":[{"href":"http:\/\/xinblog.ltd\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/xinblog.ltd\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/xinblog.ltd\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/xinblog.ltd\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=271"}],"version-history":[{"count":0,"href":"http:\/\/xinblog.ltd\/index.php?rest_route=\/wp\/v2\/posts\/271\/revisions"}],"wp:attachment":[{"href":"http:\/\/xinblog.ltd\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=271"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/xinblog.ltd\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=271"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/xinblog.ltd\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=271"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}