{"id":10006812,"date":"2022-11-04T18:58:06","date_gmt":"2022-11-04T10:58:06","guid":{"rendered":"https:\/\/ityet.com\/?p=10006812"},"modified":"2022-11-04T18:58:06","modified_gmt":"2022-11-04T10:58:06","slug":"koordinator-v0-7-%e4%b8%ba%e4%bb%bb%e5%8a%a1%e8%b0%83%e5%ba%a6%e9%a2%86%e5%9f%9f%e6%b3%a8%e5%85%a5%e6%96%b0%e6%b4%bb%e5%8a%9b","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10006812","title":{"rendered":"Koordinator v0.7: \u4e3a\u4efb\u52a1\u8c03\u5ea6\u9886\u57df\u6ce8\u5165\u65b0\u6d3b\u529b"},"content":{"rendered":"<p>\u7b80\u4ecb\uff1a\u00a0\u5728\u8fd9\u4e2a\u7248\u672c\u4e2d\u7740\u91cd\u5efa\u8bbe\u4e86\u673a\u5668\u5b66\u4e60\u3001\u5927\u6570\u636e\u573a\u666f\u9700\u8981\u7684\u4efb\u52a1\u8c03\u5ea6\u80fd\u529b\uff0c\u4f8b\u5982 Coscheduling\u3001ElasticQuota \u548c\u7cbe\u7ec6\u5316\u7684 GPU \u5171\u4eab\u8c03\u5ea6\u80fd\u529b\u3002\u5e76\u5728\u8c03\u5ea6\u95ee\u9898\u8bca\u65ad\u5206\u6790\u65b9\u9762\u5f97\u5230\u4e86\u589e\u5f3a\uff0c\u91cd\u8c03\u5ea6\u5668\u4e5f\u6781\u5927\u7684\u63d0\u5347\u4e86\u5b89\u5168\u6027\uff0c\u964d\u4f4e\u4e86\u91cd\u8c03\u5ea6\u7684\u98ce\u9669\u3002\u4f5c\u8005\uff1a\u00a0\u674e\u6d9b\uff08\u5415\u98ce\uff09\u00a0Koordinator[1]\u7ee7\u4e0a\u6b21\u00a0v0.6\u7248\u672c[2]\u53d1\u5e03\u540e\uff0c\u7ecf\u8fc7 Koordinator \u793e\u533a\u7684\u52aa\u529b\uff0c\u6211\u4eec\u8fce\u6765\u4e86\u5177\u6709\u91cd\u5927\u610f\u4e49\u7684 v0.7 \u7248\u672c\u3002\u5728\u8fd9\u4e2a\u7248\u672c\u4e2d\u7740\u91cd\u5efa\u8bbe\u4e86\u673a\u5668\u5b66\u4e60\u3001\u5927\u6570\u636e\u573a\u666f\u9700\u8981\u7684\u4efb\u52a1\u8c03\u5ea6\u80fd\u529b\uff0c\u4f8b\u5982 Coscheduling\u3001ElasticQuota \u548c\u7cbe\u7ec6\u5316\u7684 GPU \u5171\u4eab\u8c03\u5ea6\u80fd\u529b\u3002\u5e76\u5728\u8c03\u5ea6\u95ee\u9898\u8bca\u65ad\u5206\u6790\u65b9\u9762\u5f97\u5230\u4e86\u589e\u5f3a\uff0c\u91cd\u8c03\u5ea6\u5668\u4e5f\u6781\u5927\u7684\u63d0\u5347\u4e86\u5b89\u5168\u6027\uff0c\u964d\u4f4e\u4e86\u91cd\u8c03\u5ea6\u7684\u98ce\u9669\u3002\u00a0\u7248\u672c\u529f\u80fd\u7279\u6027\u89e3\u8bfb\u00a0\u4efb\u52a1\u8c03\u5ea6\u00a0Enhanced Coscheduling\u00a0Gang scheduling \u662f\u5728\u5e76\u53d1\u7cfb\u7edf\u4e2d\u5c06\u591a\u4e2a\u76f8\u5173\u8054\u7684\u8fdb\u7a0b\u8c03\u5ea6\u5230\u4e0d\u540c\u5904\u7406\u5668\u4e0a\u540c\u65f6\u8fd0\u884c\u7684\u7b56\u7565\uff0c\u5176\u6700\u4e3b\u8981\u7684\u539f\u5219\u662f\u4fdd\u8bc1\u6240\u6709\u76f8\u5173\u8054\u7684\u8fdb\u7a0b\u80fd\u591f\u540c\u65f6\u542f\u52a8\uff0c\u9632\u6b62\u90e8\u5206\u8fdb\u7a0b\u7684\u5f02\u5e38\uff0c\u5bfc\u81f4\u6574\u4e2a\u5173\u8054\u8fdb\u7a0b\u7ec4\u7684\u963b\u585e\u3002\u4f8b\u5982\u5f53\u63d0\u4ea4\u4e00\u4e2a Job \u65f6\u4f1a\u4ea7\u751f\u591a\u4e2a\u4efb\u52a1\uff0c\u8fd9\u4e9b\u4efb\u52a1\u671f\u671b\u8981\u4e48\u5168\u90e8\u8c03\u5ea6\u6210\u529f\uff0c\u8981\u4e48\u5168\u90e8\u5931\u8d25\u3002\u8fd9\u79cd\u9700\u6c42\u79f0\u4e3a All-or-Nothing\uff0c\u5bf9\u5e94\u7684\u5b9e\u73b0\u88ab\u79f0\u4f5c Gang Scheduling(or Coscheduling) \u3002\u00a0Koordinator \u5728\u542f\u52a8\u4e4b\u521d\uff0c\u671f\u671b\u652f\u6301 Kubernetes \u591a\u79cd\u5de5\u4f5c\u8d1f\u8f7d\u7684\u6df7\u90e8\u8c03\u5ea6\uff0c\u63d0\u9ad8\u5de5\u4f5c\u8d1f\u8f7d\u7684\u8fd0\u884c\u65f6\u6548\u7387\u548c\u53ef\u9760\u6027\uff0c\u5176\u4e2d\u5c31\u5305\u62ec\u4e86\u673a\u5668\u5b66\u4e60\u548c\u5927\u6570\u636e\u9886\u57df\u4e2d\u5e7f\u6cdb\u5b58\u5728\u7684\u5177\u5907 All-or-Nothing \u9700\u6c42\u7684\u4f5c\u4e1a\u8d1f\u8f7d\u3002\u4e3a\u4e86\u89e3\u51b3 All-or-Nothing \u8c03\u5ea6\u9700\u6c42\uff0cKoordinator v0.7.0 \u57fa\u4e8e\u793e\u533a\u5df2\u6709\u7684 Coscheduling \u5b9e\u73b0\u4e86 Enhanced Coscheduling\u3002\u00a0Enhanced Coscheduling \u79c9\u627f\u7740 Koordiantor \u517c\u5bb9\u793e\u533a\u7684\u539f\u5219\uff0c\u5b8c\u5168\u517c\u5bb9\u793e\u533a scheduler-plugins\/Coscheduling \u548c\u4f9d\u8d56\u7684 PodGroup CRD\u3002\u5df2\u7ecf\u4f7f\u7528 PodGroup \u7684\u7528\u6237\u53ef\u4ee5\u65e0\u7f1d\u5347\u7ea7\u5230 Koordinator\u3002\u00a0\u9664\u6b64\u4e4b\u5916\uff0cEnhanced Coscheduling \u8fd8\u5b9e\u73b0\u4e86\u5982\u4e0b\u589e\u5f3a\u80fd\u529b\uff1a\u00a01. \u652f\u6301 Strict \u548c NonStrict \u4e24\u79cd\u6a21\u5f0f\u00a0scheduler-plugins\/Coscheduling \u5728\u8c03\u5ea6\u5931\u8d25\u65f6\u4f1a Reject \u6240\u6709\u5206\u914d\u5230\u8d44\u6e90\u4f46\u5904\u4e8e Wait \u72b6\u6001\u7684 Pod\u3002\u8fd9\u79cd\u6a21\u5f0f\u6211\u4eec\u5b9a\u4e49\u4e3a Strict \u6a21\u5f0f\uff0c\u4e5f\u662f\u9ed8\u8ba4\u6a21\u5f0f\u3002\u4f46\u8fd9\u79cd\u6a21\u5f0f\u5bf9\u4e8e\u4f53\u91cf\u8f83\u5927\u7684 Job \u5e76\u4e0d\u662f\u5f88\u53cb\u597d\uff0c\u6709\u53ef\u80fd\u4f1a\u5bfc\u81f4\u8fd9\u79cd\u5927\u4f53\u91cf\u7684 Job \u62ff\u4e0d\u5230\u8d44\u6e90\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5141\u8bb8\u8c03\u5ea6\u5931\u8d25\u65f6\u4e0d\u7acb\u5373 Reject\uff0c\u5e76\u901a\u8fc7\u91cd\u8bd5\u8c03\u5ea6\u5c1d\u8bd5\u62ff\u5230\u8d44\u6e90\u6ee1\u8db3 MinMember\u3002\u8fd9\u79cd\u6a21\u5f0f\u6211\u4eec\u79f0\u4e3a NonStrict\u3002\u867d\u7136\u8fd9\u79cd\u6a21\u5f0f\u5bf9\u4e8e\u4f53\u91cf\u5927\u7684 Job \u53cb\u597d\uff0c\u53ef\u4ee5\u8ba9\u8fd9\u79cd Job \u66f4\u5feb\u7684\u8c03\u5ea6\u5b8c\u6210\uff0c\u4f46\u540c\u65f6\u4e5f\u589e\u52a0\u4e86\u8d44\u6e90\u6b7b\u9501\u7684\u98ce\u9669\u3002\u540e\u7eed Koordinator \u4f1a\u63d0\u4f9b NonStrict \u6a21\u5f0f\u4e0b\u89e3\u51b3\u6b7b\u9501\u7684\u65b9\u6848\u5b9e\u73b0\u3002\u00a0\u8981\u4f7f\u7528\u8fd9\u79cd\u6a21\u5f0f\u4e5f\u6bd4\u8f83\u7b80\u5355\uff0c\u5728 PodGroup \u6216\u8005 Pod \u4e2d\u8ffd\u52a0 annotation gang.scheduling.koordinator.sh\/mode=NonStrict \u5f00\u542f NonStrict \u6a21\u5f0f\u3002\u00a02. \u6539\u8fdb PodGroup \u8c03\u5ea6\u5931\u8d25\u7684\u5904\u7406\u673a\u5236\uff0c\u5b9e\u73b0\u66f4\u9ad8\u6548\u7684\u91cd\u8bd5\u8c03\u5ea6\u00a0\u793e\u533a scheduler-plugins\/Coscheduling \u5728\u8c03\u5ea6\u5931\u8d25\u540e\u4f1a\u5728\u4e00\u6bb5\u65f6\u95f4\u4e0d\u5728\u5c1d\u8bd5\u5904\u7406\u8be5 PodGroup\u3002\u8fd9\u4e2a\u65f6\u95f4\u4e00\u822c\u662f 20 \u79d2\u3002\u5e76\u4e14\u53ea\u652f\u6301\u6309\u7167\u63d2\u4ef6\u7ef4\u5ea6\u914d\u7f6e\uff0c\u4e0d\u652f\u6301\u7528\u6237\u6309\u4e1a\u52a1\u914d\u7f6e\u4e0d\u540c\u7684\u65f6\u95f4\u3002\u53e6\u5916\u8fd9\u79cd\u65f6\u95f4\u53c2\u6570\u4e00\u822c\u4e5f\u96be\u914d\u7f6e\u5408\u9002\u7684\u503c\u3002\u00a0\u800c\u5728 Enhanced Coscheduling \u4e2d\uff0c\u5b9e\u73b0\u4e86\u4e00\u79cd\u57fa\u4e8e ScheduleCycle \u7684\u91cd\u8bd5\u673a\u5236\u3002\u53ef\u4ee5\u7b80\u5355\u7684\u7406\u89e3\u4e3a\u4e00\u79cd\u7194\u65ad\u673a\u5236\u3002\u6bcf\u4e2a PodGroup \u4f1a\u6709\u4e00\u4e2a\u5355\u8c03\u9012\u589e\u7684\u53d8\u91cf\uff0c\u6211\u4eec\u5b9a\u4e49\u4e3a ScheduleCycle(PodGroup)\uff0c\u521d\u59cb\u503c\u4e3a 1\u3002\u8be5 PodGroup \u5173\u8054\u7684\u6240\u6709 Pod \u4e5f\u6709\u4e00\u4e2a ScheduleCycle \u53d8\u91cf\uff0c\u6211\u4eec\u5b9a\u4e49\u4e3a ScheduleCycle(Pod)\uff0c\u521d\u59cb\u503c\u4e3a ScheduleCycle(PodGroup) &#8211; 1\u3002\u6bcf\u6b21 Pod \u8c03\u5ea6\u65f6\uff0cScheduleCycle(Pod) \u7b49\u4e8e\u5f53\u524d ScheduleCycle(PodGroup)\u3002\u00a0\u5f53\u4e00\u4e2a PodGroup \u4e2d\u7684\u4e00\u4e2a Pod \u8c03\u5ea6\u5931\u8d25\u65f6\uff0c\u4f1a\u6807\u8bb0\u5f53\u524d\u7684 ScheduleCycle(PodGroup) \u65e0\u6548\uff0c\u540e\u7eed\u5176\u4ed6 Pod \u90fd\u4f1a\u8c03\u5ea6\u5931\u8d25\u3002\u5f53\u8be5 PodGroup \u4e2d\u6240\u6709 Pod ScheduleCycle(Pod) \u90fd\u4e0e\u5f53\u524d ScheduleCycle(PodGroup) \u76f8\u7b49\u65f6\uff0cScheduleCycle(PodGroup) \u5f00\u59cb\u9012\u589e\uff0c\u5141\u8bb8\u91cd\u65b0\u5f00\u59cb\u8fdb\u884c\u8c03\u5ea6\u3002\u8fd9\u79cd\u65b9\u5f0f\u53ef\u4ee5\u6709\u6548\u89c4\u907f\u57fa\u4e8e\u8fc7\u671f\u65f6\u95f4\u7684\u7f3a\u9677\uff0c\u5b8c\u5168\u53d6\u51b3\u4e8e\u8c03\u5ea6\u961f\u5217\u7684\u914d\u7f6e\u91cd\u8bd5\u8c03\u5ea6\u3002\u00a0<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042746715\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u57fa\u4e8e ScheduleCycle \u7684\u91cd\u8bd5\u673a\u5236\u00a03. \u652f\u6301\u591a\u4e2a PodGroup \u4e3a\u4e00\u7ec4\u5b8c\u6210 Gang Scheduling\u00a0\u4e00\u4e9b\u590d\u6742\u7684 Job \u6709\u591a\u79cd\u89d2\u8272\uff0c\u6bcf\u4e2a\u89d2\u8272\u7ba1\u7406\u4e00\u6279\u4efb\u52a1\uff0c\u6bcf\u4e2a\u89d2\u8272\u7684\u4efb\u52a1\u8981\u6c42\u652f\u6301 All-or-Nothing \uff0c\u6bcf\u4e2a\u89d2\u8272\u7684 MinMember \u8981\u6c42\u4e5f\u4e0d\u4e00\u6837\uff0c\u5e76\u4e14\u6bcf\u4e2a\u89d2\u8272\u4e4b\u95f4\u4e5f\u8981\u6c42 All-or-Nothing\u3002\u8fd9\u5c31\u5bfc\u81f4\u6bcf\u4e2a\u89d2\u8272\u90fd\u6709\u4e00\u4e2a\u5bf9\u5e94\u7684 PodGroup \uff0c\u5e76\u4e14\u8fd8\u8981\u6c42 PodGroup \u5373\u4f7f\u6ee1\u8db3\u4e86\u4e5f\u9700\u8981\u7b49\u5f85\u5176\u4ed6\u89d2\u8272\u7684 PodGroup \u5fc5\u987b\u6ee1\u8db3\u3002\u793e\u533a Coscheduling \u65e0\u6cd5\u6ee1\u8db3\u8fd9\u79cd\u573a\u666f\u9700\u6c42\u3002\u800c Koordinator \u5b9e\u73b0\u7684 Enhanced Coscheduling \u652f\u6301\u7528\u6237\u5728\u591a\u4e2a PodGroup \u4e2d\u589e\u52a0 anntation \u76f8\u5173\u5173\u8054\u5b9e\u73b0\uff0c\u5e76\u652f\u6301\u8de8Namespace\u3002\u4f8b\u5982\u7528\u6237\u6709 2 \u4e2a PodGroup \uff0c\u540d\u5b57\u5206\u522b\u662fPodGroupA \u548c PodGroupB\uff0c\u53ef\u4ee5\u6309\u7167\u5982\u4e0b\u4f8b\u5b50\u5173\u8054\u4e24\u4e2a PodGroup\uff1a\u00a0apiVersion: v1alpha1<br \/>kind: PodGroup<br \/>metadata:<br \/>  name: podGroupA<br \/>  namespace: default<br \/>  annotations:<\/p>\n<pre><code>gang.scheduling.koordinator.sh\/groups: [&quot;namespaceA\/podGroupA&quot;, &quot;namespaceB\/podGroupB&quot;]<\/code><\/pre>\n<p>spec:<br \/>  &#8230;\u00a04. \u652f\u6301\u8f7b\u91cf\u5316 Gang \u534f\u8bae\u00a0\u5982\u679c\u7528\u6237\u4e0d\u5e0c\u671b\u521b\u5efa PodGroup\uff0c\u8ba4\u4e3a\u521b\u5efa PodGroup \u592a\u7e41\u7410\uff0c\u90a3\u4e48\u53ef\u4ee5\u8003\u8651\u5728\u4e00\u7ec4 Pod \u4e2d\u586b\u5145\u76f8\u540c\u7684 \u00a0annotation \u00a0gang.scheduling.koordinator.sh\/name=&lt;podGroupName&gt; \u8868\u793a\u8fd9\u4e00\u7ec4 Pod \u4f7f\u7528 Coscheduling \u8c03\u5ea6\u3002\u5982\u679c\u671f\u671b\u8bbe\u7f6e minMember \uff0c\u53ef\u4ee5\u8ffd\u52a0 \u00a0Annotation gang.scheduling.koordinator.sh\/min-available=&lt;availableNum&gt;\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff1a\u00a0apiVersion: v1<br \/>kind: Pod<br \/>metadata:<br \/>  annotations:<\/p>\n<pre><code>gang.scheduling.koordinator.sh\/name: &quot;pod-group-a&quot;\ngang.scheduling.koordinator.sh\/min-available: &quot;5&quot;<\/code><\/pre>\n<p>name: demo-pod<br \/>  namespace: default<br \/>spec:<br \/>  &#8230;\u00a0ElasticQuota Scheduling\u00a0\u4e00\u5bb6\u4e2d\u5927\u578b\u516c\u53f8\u5185\u6709\u591a\u4e2a\u4ea7\u54c1\u548c\u7814\u53d1\u56e2\u961f\uff0c\u5171\u7528\u591a\u4e2a\u6bd4\u8f83\u5927\u89c4\u6a21\u7684 Kubernetes \u96c6\u7fa4\uff0c\u8fd9\u4e9b\u96c6\u7fa4\u5185\u542b\u6709\u7684\u5927\u91cf CPU\/Memory\/Disk \u7b49\u8d44\u6e90\u88ab\u8d44\u6e90\u8fd0\u8425\u56e2\u961f\u7edf\u4e00\u7ba1\u7406\u3002\u8fd0\u8425\u56e2\u961f\u5f80\u5f80\u5728\u91c7\u8d2d\u8d44\u6e90\u524d\uff0c\u901a\u8fc7\u989d\u5ea6\u9884\u7b97\u7684\u673a\u5236\u8ba9\u516c\u53f8\u5185\u6bcf\u4e2a\u56e2\u961f\u6839\u636e\u81ea\u8eab\u7684\u9700\u6c42\u63d0\u4ea4\u989d\u5ea6\u9884\u7b97\u3002\u4e1a\u52a1\u56e2\u961f\u6b64\u65f6\u4e00\u822c\u6839\u636e\u4e1a\u52a1\u5f53\u524d\u548c\u5bf9\u672a\u6765\u7684\u9884\u671f\u505a\u597d\u989d\u5ea6\u9884\u7b97\u3002\u6700\u7406\u60f3\u7684\u60c5\u51b5\u662f\u6bcf\u4e00\u4efd\u989d\u5ea6\u90fd\u80fd\u591f\u88ab\u4f7f\u7528\uff0c\u4f46\u73b0\u5b9e\u544a\u8bc9\u6211\u4eec\u8fd9\u662f\u4e0d\u73b0\u5b9e\u7684\u3002\u5f80\u5f80\u51fa\u73b0\u7684\u95ee\u9898\u662f\uff1a\u00a0\u56e2\u961f A \u9ad8\u4f30\u4e86\u4e1a\u52a1\u7684\u53d1\u5c55\u901f\u5ea6\uff0c\u7533\u8bf7\u4e86\u592a\u591a\u7684\u989d\u5ea6\u7528\u4e0d\u5b8c\u56e2\u961f B \u4f4e\u4f30\u4e86\u4e1a\u52a1\u7684\u53d1\u5c55\u901f\u5ea6\uff0c\u7533\u8bf7\u7684\u989d\u5ea6\u4e0d\u591f\u7528\u56e2\u961f C \u5b89\u6392\u4e86\u4e00\u573a\u6d3b\u52a8\uff0c\u624b\u4e0a\u7684\u989d\u5ea6\u4e0d\u591f\u591a\u4e86\uff0c\u4f46\u662f\u6d3b\u52a8\u53ea\u6301\u7eed\u51e0\u5468\uff0c\u7533\u8bf7\u592a\u591a\u989d\u5ea6\u548c\u8d44\u6e90\u4e5f\u4f1a\u6d6a\u8d39\u6389\u3002\u56e2\u961f D \u4e0b\u9762\u8fd8\u6709\u5404\u4e2a\u5b50\u56e2\u961f\u548c\u4e1a\u52a1\uff0c\u6bcf\u4e2a\u5b50\u56e2\u961f\u5185\u4e5f\u4f1a\u51fa\u73b0\u7c7b\u4f3c A B C \u4e09\u4e2a\u56e2\u961f\u7684\u60c5\u51b5\uff0c\u800c\u4e14\u5176\u4e2d\u6709\u4e9b\u56e2\u961f\u7684\u4e1a\u52a1\u4e34\u65f6\u7a81\u53d1\u9700\u8981\u63d0\u4ea4\u4e00\u4e9b\u8ba1\u7b97\u4efb\u52a1\u8981\u4ea4\u4e2a\u5ba2\u6237\uff0c\u4f46\u662f\u6ca1\u6709\u989d\u5ea6\u4e86\uff0c\u8d70\u989d\u5ea6\u9884\u7b97\u5ba1\u6279\u4e5f\u4e0d\u591f\u4e86\u3002&#8230;&#8230;\u00a0\u00a0\u4ee5\u4e0a\u5927\u5bb6\u65e5\u5e38\u7ecf\u5e38\u9047\u5230\u7684\u573a\u666f\uff0c\u5728\u6df7\u90e8\u573a\u666f\u3001\u5927\u6570\u636e\u573a\u666f\uff0c\u4e34\u65f6\u6027\u7a81\u53d1\u9700\u6c42\u53c8\u662f\u65f6\u5e38\u51fa\u73b0\u7684\uff0c\u8fd9\u4e9b\u8d44\u6e90\u7684\u9700\u6c42\u90fd\u7ed9\u989d\u5ea6\u7ba1\u7406\u5de5\u4f5c\u5e26\u6765\u4e86\u6781\u5927\u7684\u6311\u6218\u3002\u505a\u597d\u989d\u5ea6\u7ba1\u7406\u5de5\u4f5c\uff0c\u4e00\u65b9\u9762\u907f\u514d\u8fc7\u5ea6\u91c7\u8d2d\u8d44\u6e90\u964d\u4f4e\u6210\u672c\uff0c\u53c8\u8981\u5728\u4e34\u65f6\u9700\u8981\u989d\u5ea6\u65f6\u4e0d\u91c7\u8d2d\u8d44\u6e90\u6216\u8005\u5c3d\u91cf\u5c11\u7684\u91c7\u8d2d\u8d44\u6e90\uff1b\u53e6\u4e00\u65b9\u9762\u4e0d\u80fd\u56e0\u4e3a\u989d\u5ea6\u95ee\u9898\u9650\u5236\u8d44\u6e90\u4f7f\u7528\u7387\uff0c\u989d\u5ea6\u7ba1\u7406\u4e0d\u597d\u5c31\u4f1a\u5bfc\u81f4\u5373\u4f7f\u6709\u6bd4\u8f83\u597d\u7684\u6280\u672f\u5e2e\u52a9\u590d\u7528\u8d44\u6e90\uff0c\u4e5f\u65e0\u6cd5\u53d1\u6325\u5176\u4ef7\u503c\u3002\u603b\u4e4b\uff0c\u989d\u5ea6\u7ba1\u7406\u5de5\u4f5c\u662f\u5e7f\u5927\u516c\u53f8\u6216\u7ec4\u7ec7\u9700\u957f\u671f\u9762\u5bf9\u4e14\u5fc5\u987b\u9762\u5bf9\u7684\u95ee\u9898\u3002\u00a0Kubernetes ResourceQuota \u53ef\u4ee5\u89e3\u51b3\u989d\u5ea6\u7ba1\u7406\u7684\u90e8\u5206\u95ee\u9898\u3002\u539f\u751f Kubernetes ResourceQuota API \u7528\u4e8e\u6307\u5b9a\u6bcf\u4e2a Namespace \u7684\u6700\u5927\u8d44\u6e90\u989d\u5ea6\u91cf\uff0c\u5e76\u901a\u8fc7 admission \u673a\u5236\u5b8c\u6210\u51c6\u5165\u68c0\u67e5\u3002\u5982\u679c Namespace \u5f53\u524d\u8d44\u6e90\u5206\u914d\u603b\u91cf\u8d85\u8fc7 ResourceQuota \u6307\u5b9a\u7684\u914d\u989d\uff0c\u5219\u62d2\u7edd\u521b\u5efa Pod\u3002Kubernetes ResourceQuota \u8bbe\u8ba1\u6709\u4e00\u4e2a\u5c40\u9650\u6027\uff1aQuota \u00a0\u7528\u91cf\u662f\u6309\u7167 Pod Requests \u805a\u5408\u7684\u3002\u867d\u7136\u8fd9\u79cd\u673a\u5236\u53ef\u4ee5\u4fdd\u8bc1\u5b9e\u9645\u7684\u8d44\u6e90\u6d88\u8017\u6c38\u8fdc\u4e0d\u4f1a\u8d85\u8fc7 ResourceQuota \u7684\u9650\u5236\uff0c\u4f46\u5b83\u53ef\u80fd\u4f1a\u5bfc\u81f4\u8d44\u6e90\u5229\u7528\u7387\u4f4e\uff0c\u56e0\u4e3a\u4e00\u4e9b Pod \u53ef\u80fd\u5df2\u7ecf\u7533\u8bf7\u4e86\u8d44\u6e90\u4f46\u672a\u80fd\u8c03\u5ea6\u3002\u00a0Kuberenetes Scheduler-Sig \u540e\u6765\u7ed9\u51fa\u4e86\u4e00\u4e2a\u501f\u9274 Yarn Capacity Scheduling\uff0c\u79f0\u4f5c ElasticQuota \u7684\u8bbe\u8ba1\u65b9\u6848\u5e76\u7ed9\u51fa\u4e86\u5177\u4f53\u7684\u5b9e\u73b0\u3002\u5141\u8bb8\u7528\u6237\u8bbe\u7f6e max \u548c min\uff1a\u00a0max \u8868\u793a\u7528\u6237\u53ef\u4ee5\u6d88\u8d39\u7684\u8d44\u6e90\u4e0a\u9650min \u8868\u793a\u9700\u8981\u4fdd\u969c\u7528\u6237\u5b9e\u73b0\u57fa\u672c\u529f\u80fd\/\u6027\u80fd\u6240\u9700\u8981\u7684\u6700\u5c0f\u8d44\u6e90\u91cf\u00a0\u00a0\u00a0\u901a\u8fc7\u8fd9\u4e24\u4e2a\u53c2\u6570\u53ef\u4ee5\u5e2e\u52a9\u7528\u6237\u5b9e\u73b0\u5982\u4e0b\u7684\u9700\u6c42\uff1a\u00a0\u7528\u6237\u8bbe\u7f6e min &lt; max \u65f6\uff0c\u5f53\u6709\u7a81\u53d1\u8d44\u6e90\u9700\u6c42\u65f6\uff0c\u5373\u4f7f\u5f53\u524d ElasticQuota \u7684\u603b\u7528\u91cf\u8d85\u8fc7\u4e86 min\uff0c \u4f46\u53ea\u8981\u6ca1\u6709\u8fbe\u5230 max\uff0c\u90a3\u4e48\u7528\u6237\u53ef\u4ee5\u7ee7\u7eed\u521b\u5efa\u65b0\u7684 Pod \u5e94\u5bf9\u65b0\u7684\u4efb\u52a1\u8bf7\u6c42\u3002\u5f53\u7528\u6237\u9700\u8981\u66f4\u591a\u8d44\u6e90\u65f6\uff0c\u7528\u6237\u53ef\u4ee5\u4ece\u5176\u4ed6 ElasticQuota \u4e2d\u201c\u501f\u7528(borrow)\u201d \u8fd8\u6ca1\u6709\u88ab\u4f7f\u7528\u5e76\u4e14\u9700\u8981\u901a\u4fdd\u969c\u7684 min\u3002\u5f53\u4e00\u4e2a ElasticQuota \u9700\u8981\u4f7f\u7528 min \u8d44\u6e90\u65f6\uff0c\u4f1a\u901a\u8fc7\u62a2\u5360\u673a\u5236\u4ece\u5176\u4ed6\u501f\u7528\u65b9\u62a2\u56de\u6765\uff0c\u5373\u9a71\u9010\u4e00\u4e9b\u5176\u4ed6 ElasticQuota \u8d85\u8fc7 min \u7528\u91cf\u7684 Pod\u3002\u00a0\u00a0\u00a0ElasticQuota \u8fd8\u6709\u4e00\u4e9b\u5c40\u9650\u6027\uff1a\u6ca1\u6709\u5f88\u597d\u7684\u4fdd\u969c\u516c\u5e73\u6027\u3002\u5047\u5982\u540c\u4e00\u4e2a ElasticQuota \u6709\u5927\u91cf\u65b0\u5efa\u7684 Pod\uff0c\u6709\u53ef\u80fd\u4f1a\u6d88\u8017\u6240\u6709\u5176\u4ed6\u53ef\u4ee5\u88ab\u501f\u7528\u7684 Quota\uff0c\u4ece\u800c\u5bfc\u81f4\u540e\u6765\u7684 Pod \u53ef\u80fd\u62ff\u4e0d\u5230 Quota\u3002\u6b64\u65f6\u53ea\u80fd\u901a\u8fc7\u62a2\u5360\u673a\u5236\u62a2\u56de\u6765\u4e00\u4e9b Quota\u3002\u00a0\u53e6\u5916 ElasticQuota \u548c Kubernetes ResourceQuota \u90fd\u662f\u9762\u5411 Namespace \u7684\uff0c\u4e0d\u652f\u6301\u591a\u7ea7\u6811\u5f62\u7ed3\u6784\uff0c\u5bf9\u4e8e\u4e00\u4e9b\u672c\u8eab\u5177\u5907\u590d\u6742\u7ec4\u7ec7\u5173\u7cfb\u7684\u4f01\u4e1a\/\u7ec4\u7ec7\u4e0d\u80fd\u5f88\u597d\u7684\u4f7f\u7528 ElasticQuota\/Kubenretes ResourceQuota \u5b8c\u6210\u989d\u5ea6\u7ba1\u7406\u5de5\u4f5c\u3002\u00a0Koordinator \u9488\u5bf9\u8fd9\u4e9b\u989d\u5ea6\u7ba1\u7406\u95ee\u9898\uff0c\u7ed9\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u793e\u533a ElasticQuota \u5b9e\u73b0\u7684\u652f\u6301\u591a\u7ea7\u7ba1\u7406\u65b9\u5f0f\u7684\u5f39\u6027 Quota \u7ba1\u7406\u673a\u5236(multi hierarchy quota management)\u3002\u5177\u5907\u5982\u4e0b\u7279\u6027\uff1a\u00a0\u517c\u5bb9\u793e\u533a\u7684 ElasticQuota API\u3002\u7528\u6237\u53ef\u4ee5\u65e0\u7f1d\u5347\u7ea7\u5230 Koordinator\u652f\u6301\u6811\u5f62\u7ed3\u6784\u7ba1\u7406 Quota\u3002\u652f\u6301\u6309\u7167\u5171\u4eab\u6743\u91cd(shared weight)\u4fdd\u969c\u516c\u5e73\u6027\u3002\u5141\u8bb8\u7528\u6237\u8bbe\u7f6e\u662f\u5426\u5141\u8bb8\u501f\u7528 Quota \u7ed9\u5176\u4ed6\u6d88\u8d39\u5bf9\u8c61\u3002\u00a0\u00a01. Pod \u5173\u8054 ElasticQuota \u65b9\u5f0f\u00a0\u7528\u6237\u53ef\u4ee5\u975e\u5e38\u4f7f\u7528\u7684\u4f7f\u7528\u8be5\u80fd\u529b\uff0c\u53ef\u4ee5\u5b8c\u5168\u6309\u7167 ElasticQuota \u7684\u7528\u6cd5\uff0c\u5373\u6bcf\u4e2a Namespace \u8bbe\u7f6e\u4e00\u4e2a ElasticQuota \u5bf9\u8c61\u3002\u4e5f\u53ef\u4ee5\u5728 Pod \u4e2d\u8ffd\u52a0 Label \u5173\u8054 ElasticQuota\uff1a\u00a0apiVersion: v1<br \/>kind: Pod<br \/>metadata:<br \/>  labels:<\/p>\n<pre><code>quota.scheduling.koordinator.sh\/name: &quot;elastic-quota-a&quot;<\/code><\/pre>\n<p>name: demo-pod<br \/>  namespace: default<br \/>spec:<br \/>  &#8230;\u00a02. \u6811\u5f62\u7ed3\u6784\u7ba1\u7406\u673a\u5236\u548c\u4f7f\u7528\u65b9\u6cd5\u00a0\u9700\u8981\u4f7f\u7528\u6811\u5f62\u7ed3\u6784\u7ba1\u7406 Quota \u65f6\uff0c\u9700\u8981\u5728 ElasticQuota \u4e2d\u8ffd\u52a0 Label \u00a0quota.scheduling.koordinator.sh\/is-parent\u8868\u793a\u5f53\u524d ElasticQuota \u662f\u5426\u662f\u7236\u8282\u70b9\uff0cquota.scheduling.koordinator.sh\/parent\u8868\u793a\u5f53\u524d ElasticQuota \u7684\u7236\u8282\u70b9 ElasticQuota \u7684\u540d\u5b57\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff1a\u00a0<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042746716\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u6211\u4eec\u521b\u5efa\u4e00\u4e2a ElasticQuota parentA \u4f5c\u4e3a\u7236\u8282\u70b9\uff0c\u8d44\u6e90\u603b\u91cf\u4e3aCPU 100C\uff0c\u5185\u5b58 200Gi\uff0c\u4ee5\u53ca\u5b50\u8282\u70b9 childA1\u00a0apiVersion: scheduling.sigs.k8s.io\/v1alpha1<br \/>kind: ElasticQuota<br \/>metadata:<br \/>  name: parentA<br \/>  namespace: default<br \/>  labels:<\/p>\n<pre><code>quota.scheduling.koordinator.sh\/is-parent: &quot;true&quot;\nquota.scheduling.koordinator.sh\/allow-lent-resource: &quot;true&quot;<\/code><\/pre>\n<p>spec:<br \/>  max:<\/p>\n<pre><code>cpu: 100\nmemory: 200Gi<\/code><\/pre>\n<p>min:<\/p>\n<pre><code>cpu: 100\nmemory: 200Gi<\/code><\/pre>\n<hr>\n<p>apiVersion: scheduling.sigs.k8s.io\/v1alpha1<br \/>kind: ElasticQuota<br \/>metadata:<br \/>  name: childA1<br \/>  namespace: default<br \/>  labels:<\/p>\n<pre><code>quota.scheduling.koordinator.sh\/is-parent: &quot;false&quot;\nquota.scheduling.koordinator.sh\/parent: &quot;parentA&quot;\nquota.scheduling.koordinator.sh\/allow-lent-resource: &quot;true&quot;<\/code><\/pre>\n<p>spec:<br \/>  max:<\/p>\n<pre><code>cpu: 40\nmemory: 100Gi<\/code><\/pre>\n<p>min:<\/p>\n<pre><code>cpu: 20\nmemory: 40Gi\u00a0\u5728\u4f7f\u7528\u6811\u5f62\u7ed3\u6784\u7ba1\u7406 ElasticQuota \u65f6\uff0c\u6709\u4e00\u4e9b\u9700\u8981\u9075\u5faa\u7684\u7ea6\u675f\uff1a\u00a0\u9664\u4e86\u6839\u8282\u70b9\uff0c\u5176\u4ed6\u6240\u6709\u5b50\u8282\u70b9\u7684 min \u4e4b\u548c\u8981\u5c0f\u4e8e\u7236\u8282\u70b9\u7684 min\u3002\u00a0\u00a0\u4e0d\u9650\u5236\u5b50\u8282\u70b9 max\uff0c\u5141\u8bb8\u5b50\u8282\u70b9\u7684 max \u5927\u4e8e\u7236\u8282\u70b9\u7684 max\u3002\u8003\u8651\u4ee5\u4e0b\u573a\u666f\uff0c\u96c6\u7fa4\u4e2d\u6709 2 \u4e2a ElasticQuota \u5b50\u6811\uff1adev-parent \u548c production-parent\uff0c\u6bcf\u4e2a\u5b50\u6811\u90fd\u6709\u51e0\u4e2a\u5b50 ElasticQuota\u3002\u5f53 production-parent \u5fd9\u65f6\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u53ea\u964d\u4f4e dev-parent \u7684 max \u9650\u5236 \u00a0dev-parent \u6574\u9897\u5b50\u6811\u7684\u8d44\u6e90\u4f7f\u7528\u91cf\uff0c\u800c\u4e0d\u662f\u964d\u4f4e dev-parent \u5b50\u6811\u7684\u6bcf\u4e2a\u5b50 ElasticQuota \u7684 max \u9650\u5236\u7528\u91cf\u3002\u00a0\u00a0Pod \u4e0d\u80fd\u4f7f\u7528\u7236\u8282\u70b9 ElasticQuota\u3002\u5982\u679c\u653e\u5f00\u8fd9\u4e2a\u9650\u5236\uff0c\u4f1a\u5bfc\u81f4\u6574\u4e2a\u5f39\u6027 Quota \u7684\u673a\u5236\u53d8\u7684\u5f02\u5e38\u590d\u6742\uff0c\u6682\u65f6\u4e0d\u8003\u8651\u652f\u6301\u8fd9\u79cd\u573a\u666f\u3002\u00a0\u00a0\u53ea\u6709\u7236\u8282\u70b9\u53ef\u4ee5\u6302\u5b50\u8282\u70b9\uff0c\u4e0d\u5141\u8bb8\u5b50\u8282\u70b9\u6302\u5b50\u8282\u70b9\u3002\u00a0\u00a0\u6682\u65f6\u4e0d\u5141\u8bb8\u6539\u53d8 ElasticQuota \u7684 quota.scheduling.koordinator.sh\/is-parent \u5c5e\u6027\u00a0\u00a0\u6211\u4eec\u5c06\u5728\u4e0b\u4e2a\u7248\u672c\u4e2d\u901a\u8fc7 webhook \u673a\u5236\u5b9e\u73b0\u8fd9\u4e9b\u7ea6\u675f\u3002\u00a03. \u516c\u5e73\u6027\u4fdd\u969c\u673a\u5236\u00a0Koordinator ElasticQuota \u00a0\u652f\u6301\u6309\u7167\u5171\u4eab\u6743\u91cd(shared weight)\u4fdd\u969c\u516c\u5e73\u6027\u3002shared-weight \u8868\u793a\u4e00\u4e2a ElasticQuota \u5bf9\u8c61 \u7684\u7ade\u4e89\u529b\uff0c\u9ed8\u8ba4\u7b49\u4e8e ElasticQuota Max\u3002\u901a\u8fc7\u516c\u5e73\u6027\u4fdd\u969c\u673a\u5236\uff0c\u4f1a\u7ed9\u6240\u6709 min &lt; max \u7684 ElasticQuota \u5206\u914d\u5b9e\u9645\u53ef\u7528\u7684\u8d44\u6e90\u91cf\uff0c\u8be5\u8d44\u6e90\u91cf\u7528 runtime\u8868\u793a\uff0c\u5e76\u4fdd\u969c runtime \u5728min \u4e0e max \u4e4b\u95f4\uff0c\u5373 max &gt;= runtime &gt;= min\u3002\u00a0\u5f53\u4e00\u4e2a ElasticQuota \u5bf9\u8c61\u5173\u8054\u7684\u6240\u6709 Pod \u7684\u8d44\u6e90\u8bf7\u6c42\u91cf\uff08\u6211\u4eec\u5b9a\u4e49\u4e3a request\uff09\u5927\u4e8e min \u65f6\uff0c\u4f1a\u501f\u7528\u5176\u4ed6 ElasticQuota \u5bf9\u8c61\u4e2d min \u672a\u5206\u914d\u7684\u90e8\u5206\u3002\u88ab\u501f\u51fa\u53bb\u7684 Quota \u9700\u8981\u4f7f\u7528\u65f6\uff0c\u4f1a\u518d\u6b21\u901a\u8fc7\u8be5\u516c\u5e73\u6027\u4fdd\u969c\u673a\u5236\u62ff\u56de\u6765\u3002\u00a0\u53e6\u5916\u8fd8\u6709\u4e00\u79cd\u65e5\u5e38\u751f\u4ea7\u65f6\u4f1a\u9047\u5230\u7684\u60c5\u51b5\uff1a\u5373\u96c6\u7fa4\u5185\u8d44\u6e90\u603b\u91cf\u4f1a\u968f\u7740\u8282\u70b9\u6545\u969c\u3001\u8d44\u6e90\u8fd0\u8425\u7b49\u539f\u56e0\u964d\u4f4e\uff0c\u5bfc\u81f4\u6240\u6709 ElasticQuota \u7684 min \u4e4b\u548c\u5927\u4e8e\u8d44\u6e90\u603b\u91cf\u3002\u5f53\u51fa\u73b0\u8fd9\u79cd\u60c5\u51b5\u65f6\uff0c\u6211\u4eec\u65e0\u6cd5\u786e\u4fdd min \u7684\u8d44\u6e90\u9700\u6c42\u3002\u6b64\u65f6\u6211\u4eec\u4f1a\u6309\u7167\u4e00\u5b9a\u7684\u6bd4\u4f8b\u8c03\u6574\u5404\u4e2a ElasticQuota \u7684 min\uff0c\u786e\u4fdd\u6240\u6709 min \u4e4b\u548c\u5c0f\u4e8e\u6216\u8005\u7b49\u4e8e\u5f53\u524d\u5b9e\u9645\u7684\u8d44\u6e90\u603b\u91cf\u3002\u00a04. \u62a2\u5360\u673a\u5236\u00a0Koordinator ElasticQuota \u673a\u5236\u5728\u8c03\u5ea6\u9636\u6bb5\u5982\u679c\u53d1\u73b0 Quota \u4e0d\u8db3\uff0c\u4f1a\u8fdb\u5165\u62a2\u5360\u9636\u6bb5\uff0c\u6309\u7167\u4f18\u5148\u7ea7\u6392\u5e8f\uff0c\u62a2\u5360\u5c5e\u4e8e\u540c\u4e00\u4e2aElasticQuota \u5185\u7684 \u4f4e\u4f18\u5148\u7ea7 Pod\u3002\u540c\u65f6\uff0c\u6211\u4eec\u4e0d\u652f\u6301\u8de8 ElasticQuota \u62a2\u5360\u5176\u4ed6 Pod\u3002\u4f46\u662f\u6211\u4eec\u4e5f\u63d0\u4f9b\u4e86\u53e6\u5916\u7684\u673a\u5236\u652f\u6301\u4ece\u501f\u7528 Quota \u7684 ElasticQuota \u62a2\u56de\u3002\u00a0\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5728\u96c6\u7fa4\u4e2d\uff0c\u6709\u4e24\u4e2a ElasticQuota\uff0cElasticQuota A {min = 50, max = 100}\uff0c ElasticQuota B {min = 50, \u00a0max = 100}\u3002\u7528\u6237\u5728\u4e0a\u5348 10 \u70b9\u4f7f\u7528 ElasticQuota A \u63d0\u4ea4\u4e86\u4e00\u4e2a Job\uff0c Request = 100 \uff0c\u6b64\u65f6\u56e0\u4e3a ElasticQuota B \u65e0\u4eba\u4f7f\u7528\uff0cElasticQuota A \u80fd\u4ece B \u624b\u91cc\u501f\u7528 50 \u4e2a Quota\uff0c\u6ee1\u8db3\u4e86 Request = 100\uff0c \u5e76\u4e14\u6b64\u65f6 Used = 100\u3002\u5728 11 \u70b9\u949f\u65f6\uff0c\u53e6\u4e00\u4e2a\u7528\u6237\u5f00\u59cb\u4f7f\u7528 ElasticQuota B \u63d0\u4ea4 Job\uff0cRequest = 100\uff0c\u56e0\u4e3a ElasticQuota B \u7684 min = 50\uff0c\u662f\u5fc5\u987b\u4fdd\u969c\u7684\uff0c\u901a\u8fc7\u516c\u5e73\u6027\u4fdd\u969c\u673a\u5236\uff0c\u6b64\u65f6 A \u548c B \u7684 runtime \u5747\u4e3a 50\u3002\u90a3\u4e48\u6b64\u65f6\u5bf9\u4e8e ElasticQuota A \uff0cUsed = 100 \u662f\u5927\u4e8e\u5f53\u524d runtime = 50 \u7684\uff0c\u56e0\u6b64\u6211\u4eec\u4f1a\u63d0\u4f9b\u4e00\u4e2a Controller\uff0c\u9a71\u9010\u6389\u4e00\u90e8\u5206 Pod \uff0c\u4f7f\u5f97\u5f53\u524d ElasticQuota A \u7684 Used \u964d\u4f4e\u5230 runtime \u76f8\u7b49\u7684\u6c34\u4f4d\u3002\u00a0\u7cbe\u7ec6\u5316\u8d44\u6e90\u8c03\u5ea6\u00a0Device Share Scheduling\u00a0\u673a\u5668\u5b66\u4e60\u9886\u57df\u91cc\u4f9d\u9760\u5927\u91cf\u5f3a\u5927\u7b97\u529b\u6027\u80fd\u7684 GPU \u8bbe\u5907\u5b8c\u6210\u6a21\u578b\u8bad\u7ec3\uff0c\u4f46\u662f GPU \u81ea\u8eab\u4ef7\u683c\u5341\u5206\u6602\u8d35\u3002\u5982\u4f55\u66f4\u597d\u5730\u5229\u7528 GPU \u8bbe\u5907\uff0c\u53d1\u6325 GPU \u7684\u4ef7\u503c\uff0c\u964d\u4f4e\u6210\u672c\uff0c\u662f\u4e00\u4e2a\u4e9f\u5f85\u89e3\u51b3\u7684\u95ee\u9898\u3002Kubernetes \u793e\u533a\u73b0\u6709\u7684 GPU \u5206\u914d\u673a\u5236\u4e2d\uff0cGPU \u662f\u7531 kubelet \u5206\u914d\u7684\uff0c\u5e76\u53ea\u652f\u6301\u5206\u914d\u4e00\u4e2a\u6216\u591a\u4e2a\u5b8c\u6574\u7684 GPU \u5b9e\u4f8b\u3002\u8fd9\u79cd\u65b9\u6cd5\u7b80\u5355\u53ef\u9760\uff0c\u4f46\u7c7b\u4f3c\u4e8e CPU \u548c Memory\uff0cGPU \u5e76\u4e0d\u662f\u4e00\u76f4\u5904\u4e8e\u9ad8\u5229\u7528\u7387\u6c34\u4f4d\uff0c\u540c\u6837\u5b58\u5728\u8d44\u6e90\u6d6a\u8d39\u7684\u95ee\u9898\u3002\u56e0\u6b64\uff0cKoordinator \u5e0c\u671b\u652f\u6301\u591a\u5de5\u4f5c\u8d1f\u8f7d\u5171\u4eab\u4f7f\u7528 GPU \u8bbe\u5907\u4ee5\u8282\u7701\u6210\u672c\u3002\u6b64\u5916\uff0cGPU \u6709\u5176\u7279\u6b8a\u6027\u3002\u6bd4\u5982\u4e0b\u9762\u7684 NVIDIA GPU \u652f\u6301\u7684 NVLink \u548c\u8d85\u5356\u573a\u666f\uff0c\u90fd\u9700\u8981\u901a\u8fc7\u8c03\u5ea6\u5668\u8fdb\u884c\u4e2d\u592e\u51b3\u7b56\uff0c\u4ee5\u83b7\u5f97\u5168\u5c40\u6700\u4f18\u7684\u5206\u914d\u7ed3\u679c\u3002\u00a0<\/code><\/pre>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042746717\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u4ece\u56fe\u4e2d\u6211\u4eec\u53ef\u4ee5\u53d1\u73b0\uff0c\u867d\u7136\u8be5\u8282\u70b9\u6709 8 \u4e2a GPU \u5b9e\u4f8b\uff0c\u578b\u53f7\u4e3a A100\/V100\uff0c\u4f46 GPU \u5b9e\u4f8b\u4e4b\u95f4\u7684\u6570\u636e\u4f20\u8f93\u901f\u5ea6\u662f\u4e0d\u540c\u7684\u3002\u5f53\u4e00\u4e2a Pod \u9700\u8981\u591a\u4e2a GPU \u5b9e\u4f8b\u65f6\uff0c\u6211\u4eec\u53ef\u4ee5\u4e3a Pod \u5206\u914d\u5177\u6709\u6700\u5927\u6570\u636e\u4f20\u8f93\u901f\u5ea6\u7ec4\u5408\u5173\u7cfb\u7684 GPU \u5b9e\u4f8b\u3002\u6b64\u5916\uff0c\u5f53\u6211\u4eec\u5e0c\u671b\u4e00\u7ec4 Pod \u4e2d\u7684 GPU \u5b9e\u4f8b\u5177\u6709\u6700\u5927\u6570\u636e\u4f20\u8f93\u901f\u5ea6\u7ec4\u5408\u5173\u7cfb\u65f6\uff0c\u8c03\u5ea6\u5668\u5e94\u8be5\u5c06\u6700\u4f73 GPU \u5b9e\u4f8b\u6279\u91cf\u5206\u914d\u7ed9\u8fd9\u4e9b Pod\uff0c\u5e76\u5c06\u5b83\u4eec\u5206\u914d\u5230\u540c\u4e00\u4e2a\u8282\u70b9\u3002\u00a01. GPU \u8d44\u6e90\u534f\u8bae\u00a0Koordinator \u517c\u5bb9\u793e\u533a\u5df2\u6709\u7684 nvidia.com\/gpu \u8d44\u6e90\u534f\u8bae\uff0c\u5e76\u4e14\u8fd8\u81ea\u5b9a\u4e49\u4e86\u6269\u5c55\u8d44\u6e90\u534f\u8bae\uff0c\u652f\u6301\u7528\u6237\u66f4\u7ec6\u7c92\u5ea6\u7684\u5206\u914d GPU \u8d44\u6e90\u3002\u00a0kubernetes.io\/gpu-core \u4ee3\u8868 GPU \u7684\u8ba1\u7b97\u80fd\u529b\u3002\u4e0e Kuberetes MilliCPU \u7c7b\u4f3c\uff0c\u6211\u4eec\u5c06 GPU \u7684\u603b\u7b97\u529b\u62bd\u8c61\u4e3a 100\uff0c\u7528\u6237\u53ef\u4ee5\u6839\u636e\u9700\u8981\u7533\u8bf7\u76f8\u5e94\u6570\u91cf\u7684 GPU \u7b97\u529b\u3002kubernetes.io\/gpu-memory \u8868\u793a GPU \u7684\u5185\u5b58\u5bb9\u91cf\uff0c\u4ee5\u5b57\u8282\u4e3a\u5355\u4f4d\u3002kubernetes.io\/gpu-memory-ratio \u4ee3\u8868 GPU \u5185\u5b58\u7684\u767e\u5206\u6bd4\u3002\u00a0\u00a0\u00a0\u5047\u8bbe\u4e00\u4e2a\u8282\u70b9\u6709 4 \u4e2a GPU \u8bbe\u5907\u5b9e\u4f8b\uff0c\u6bcf\u4e2a GPU \u8bbe\u5907\u5b9e\u4f8b\u6709 8Gi \u663e\u5b58\u3002\u7528\u6237\u5982\u679c\u671f\u671b\u7533\u8bf7\u4e00\u4e2a\u5b8c\u6574\u7684 GPU \u5b9e\u4f8b\uff0c\u9664\u4e86\u4f7f\u7528 nvidia.com\/gpu \u4e4b\u5916\uff0c\u8fd8\u53ef\u4ee5\u6309\u7167\u5982\u4e0b\u65b9\u5f0f\u7533\u8bf7\uff1a\u00a0apiVersion: v1<br \/>kind: Pod<br \/>metadata:<br \/>  name: demo-pod<br \/>  namespace: default<br \/>spec:<br \/>  containers:<\/p>\n<ul>\n<li>\n<p>name: main<br \/>resources:<br \/>  limits:<\/p>\n<pre><code>kubernetes.io\/gpu-core: 100\nkubernetes.io\/gpu-memory: &quot;8Gi&quot;<\/code><\/pre>\n<p>requests:<\/p>\n<pre><code>kubernetes.io\/gpu-core: 100\nkubernetes.io\/gpu-memory: &quot;8Gi&quot;\u00a0\u5982\u679c\u671f\u671b\u53ea\u4f7f\u7528\u4e00\u4e2a GPU \u5b9e\u4f8b\u4e00\u534a\u7684\u8d44\u6e90\uff0c\u53ef\u4ee5\u6309\u7167\u5982\u4e0b\u65b9\u5f0f\u7533\u8bf7\uff1a\u00a0apiVersion: v1<\/code><\/pre>\n<\/li>\n<\/ul>\n<p>kind: Pod<br \/>metadata:<br \/>  name: demo-pod<br \/>  namespace: default<br \/>spec:<br \/>  containers:<\/p>\n<ul>\n<li>\n<p>name: main<br \/>resources:<br \/>  limits:<\/p>\n<pre><code>kubernetes.io\/gpu-core: 50\nkubernetes.io\/gpu-memory: &quot;4Gi&quot;<\/code><\/pre>\n<p>requests:<\/p>\n<pre><code>kubernetes.io\/gpu-core: 50\nkubernetes.io\/gpu-memory: &quot;4Gi&quot;\u00a02. \u8bbe\u5907\u4fe1\u606f\u548c\u8bbe\u5907\u5bb9\u91cf\u4e0a\u62a5\u00a0\u5728 Koordinator v0.7.0 \u7248\u672c\u4e2d\uff0c\u5355\u673a\u4fa7 koordlet \u5b89\u88c5\u540e\u4f1a\u81ea\u52a8\u8bc6\u522b\u8282\u70b9\u4e0a\u662f\u5426\u542b\u6709 GPU \u8bbe\u5907\uff0c\u5982\u679c\u5b58\u5728\u7684\u8bdd\uff0c\u4f1a\u4e0a\u62a5\u8fd9\u4e9b GPU \u8bbe\u5907\u7684 Minor ID\u3001 UUID\u3001\u7b97\u529b\u548c\u663e\u5b58\u5927\u5c0f\u5230\u4e00\u4e2a\u7c7b\u578b\u4e3a Device CRD \u4e2d\u3002\u6bcf\u4e2a\u8282\u70b9\u5bf9\u5e94\u4e00\u4e2a Device CRD \u5b9e\u4f8b\u3002Device CRD \u4e0d\u4ec5\u652f\u6301\u63cf\u8ff0 GPU\uff0c\u8fd8\u652f\u6301\u7c7b\u4f3c\u4e8e FPGA\/RDMA \u7b49\u8bbe\u5907\u7c7b\u578b\uff0c\u76ee\u524d v0.7.0 \u7248\u672c\u53ea\u652f\u6301 GPU\uff0c \u6682\u672a\u652f\u6301\u8fd9\u4e9b\u8bbe\u5907\u7c7b\u578b\u3002\u00a0Device CRD \u4f1a\u88ab koord-manager \u5185\u7684 NodeResource controller \u548c koord-scheduler \u6d88\u8d39\u3002NodeResource controller \u4f1a\u6839\u636e Device CRD \u4e2d\u63cf\u8ff0\u7684\u4fe1\u606f\uff0c\u6362\u7b97\u6210 Koordinator \u652f\u6301\u7684\u8d44\u6e90\u534f\u8bae kubernetes.io\/gpu-core,kubernetes.io\/gpu-memory \u66f4\u65b0\u5230 \u00a0Node.Status.Allocatable \u548c Node.Status.Capacity \u5b57\u6bb5\uff0c\u5e2e\u52a9\u8c03\u5ea6\u5668\u548c kubelet \u5b8c\u6210\u8d44\u6e90\u8c03\u5ea6\u3002gpu-core \u8868\u793a GPU \u8bbe\u5907\u5b9e\u4f8b\u7684\u7b97\u529b\uff0c\u4e00\u4e2a\u5b9e\u4f8b\u7684\u5b8c\u6574\u7b97\u529b\u4e3a100\u3002\u5047\u8bbe\u4e00\u4e2a\u8282\u70b9\u6709 8 \u4e2a GPU \u8bbe\u5907\u5b9e\u4f8b\uff0c\u90a3\u4e48\u8282\u70b9\u7684 gpu-core \u5bb9\u91cf\u4e3a 8 * 100 = 800\uff1bgpu-memory \u8868\u793a GPU \u8bbe\u5907\u5b9e\u4f8b\u7684\u663e\u5b58\u5927\u5c0f\uff0c\u5355\u4f4d\u4e3a\u5b57\u8282\uff0c\u540c\u6837\u7684\u8282\u70b9\u53ef\u4ee5\u5206\u914d\u7684\u663e\u5b58\u603b\u91cf\u4e3a \u8bbe\u5907\u6570\u91cf * \u6bcf\u4e2a\u5b9e\u4f8b\u7684\u5355\u4f4d\u5bb9\u91cf\uff0c\u4f8b\u5982\u4e00\u4e2a GPU \u8bbe\u5907\u7684\u663e\u5b58\u662f 8G\uff0c\u8282\u70b9\u4e0a\u6709 8 \u4e2a GPU \u5b9e\u4f8b\uff0c\u603b\u91cf\u4e3a 8 * 8G = 64G\u3002\u00a0apiVersion: v1<\/code><\/pre>\n<\/li>\n<\/ul>\n<p>kind: Node<br \/>metadata:<br \/>  name: node-a<br \/>status:<br \/>  capacity:<\/p>\n<pre><code>koordinator.sh\/gpu-core: 800\nkoordinator.sh\/gpu-memory: &quot;64Gi&quot;\nkoordinator.sh\/gpu-memory-ratio: 800<\/code><\/pre>\n<p>allocatable:<\/p>\n<pre><code>koordinator.sh\/gpu-core: 800\nkoordinator.sh\/gpu-memory: &quot;64Gi&quot;\nkoordinator.sh\/gpu-memory-ratio: 800\u00a03. \u4e2d\u5fc3\u8c03\u5ea6\u5206\u914d\u8bbe\u5907\u8d44\u6e90\u00a0Kuberetes \u793e\u533a\u539f\u751f\u63d0\u4f9b\u7684\u8bbe\u5907\u8c03\u5ea6\u673a\u5236\u4e2d\uff0c\u8c03\u5ea6\u5668\u53ea\u8d1f\u8d23\u6821\u9a8c\u8bbe\u5907\u5bb9\u91cf\u662f\u5426\u6ee1\u8db3 Pod\uff0c\u5bf9\u4e8e\u4e00\u4e9b\u7b80\u5355\u7684\u8bbe\u5907\u7c7b\u578b\u662f\u8db3\u591f\u7684\uff0c\u4f46\u662f\u5f53\u9700\u8981\u66f4\u7ec6\u7c92\u5ea6\u5206\u914d GPU \u65f6\uff0c\u9700\u8981\u4e2d\u5fc3\u8c03\u5ea6\u5668\u7ed9\u4e88\u652f\u6301\u624d\u80fd\u5b9e\u73b0\u5168\u5c40\u6700\u4f18\u3002\u00a0Koordinator \u8c03\u5ea6\u5668 koord-scheduler \u65b0\u589e\u4e86\u8c03\u5ea6\u63d2\u4ef6 DeviceShare\uff0c\u8d1f\u8d23\u7cbe\u7ec6\u5ea6\u8bbe\u5907\u8d44\u6e90\u8c03\u5ea6\u3002DeviceShare \u63d2\u4ef6\u6d88\u8d39 Device CRD\uff0c\u8bb0\u5f55\u6bcf\u4e2a\u8282\u70b9\u53ef\u4ee5\u5206\u914d\u7684\u8bbe\u5907\u4fe1\u606f\u3002DeviceShare \u5728\u8c03\u5ea6\u65f6\uff0c\u4f1a\u628a Pod \u7684GPU\u8d44\u6e90\u8bf7\u6c42\u8f6c\u6362\u4e3a Koordinator \u7684\u8d44\u6e90\u534f\u8bae\uff0c\u5e76\u8fc7\u6ee4\u6bcf\u4e2a\u8282\u70b9\u7684\u672a\u5206\u914d\u7684 GPU \u8bbe\u5907\u5b9e\u4f8b\u3002\u786e\u4fdd\u6709\u8d44\u6e90\u53ef\u7528\u540e\uff0c\u5728 Reserve \u9636\u6bb5\u66f4\u65b0\u5185\u90e8\u72b6\u6001\uff0c\u5e76\u5728 PreBind \u9636\u6bb5\u66f4\u65b0 Pod Annotation\uff0c\u8bb0\u5f55\u5f53\u524d Pod \u5e94\u8be5\u4f7f\u7528\u54ea\u4e9b GPU \u8bbe\u5907\u3002\u00a0DeviceShare \u5c06\u5728\u540e\u7eed\u7248\u672c\u652f\u6301 Binpacking \u00a0\u548c Spread \u7b56\u7565\uff0c\u5b9e\u73b0\u66f4\u597d\u7684\u8bbe\u5907\u8d44\u6e90\u8c03\u5ea6\u80fd\u529b\u3002\u00a04. \u5355\u673a\u4fa7\u7cbe\u51c6\u7ed1\u5b9a\u8bbe\u5907\u4fe1\u606f\u00a0Kubernetes \u793e\u533a\u5728 kubelet \u4e2d\u63d0\u4f9b\u4e86 DevicePlugin \u673a\u5236\uff0c\u652f\u6301\u8bbe\u5907\u5382\u5546\u5728 kubelet \u5206\u914d\u597d\u8bbe\u5907\u540e\u6709\u673a\u4f1a\u83b7\u5f97\u8bbe\u5907\u4fe1\u606f\uff0c\u5e76\u586b\u5145\u5230\u73af\u5883\u53d8\u91cf\u6216\u8005\u66f4\u65b0\u6302\u8f7d\u8def\u5f84\u3002\u4f46\u662f\u4e0d\u80fd\u652f\u6301 \u4e2d\u5fc3\u5316\u7684 GPU \u7cbe\u7ec6\u5316\u8c03\u5ea6\u573a\u666f\u3002\u00a0\u9488\u5bf9\u8fd9\u4e2a\u95ee\u9898\uff0c Koordinator \u6269\u5c55\u4e86 koord-runtime-proxy \uff0c\u652f\u6301\u5728 kubelet \u521b\u5efa\u5bb9\u5668\u65f6\u66f4\u65b0\u73af\u5883\u53d8\u91cf\uff0c\u6ce8\u5165\u8c03\u5ea6\u5668\u5206\u914d\u7684 GPU \u8bbe\u5907\u4fe1\u606f\u3002\u00a0<\/code><\/pre>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042746718\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u8c03\u5ea6\u5668\u8bca\u65ad\u5206\u6790\u00a0\u5927\u5bb6\u5728\u4f7f\u7528 Kubernetes \u65f6\u7ecf\u5e38\u4f1a\u9047\u5230\u4e00\u4e9b\u8c03\u5ea6\u76f8\u5173\u7684\u95ee\u9898\uff1a\u00a0\u6211\u8fd9\u4e2a Pod \u4e3a\u4ec0\u4e48\u4e0d\u80fd\u8c03\u5ea6\uff1f\u8fd9\u4e2a Pod \u4e3a\u4ec0\u4e48\u4f1a\u8c03\u5ea6\u5230\u8fd9\u4e2a\u8282\u70b9\uff0c\u4e0d\u662f\u5e94\u8be5\u88ab\u53e6\u4e00\u4e2a\u6253\u5206\u63d2\u4ef6\u5f71\u54cd\u5230\u4e48\uff1f\u6211\u65b0\u5f00\u53d1\u4e86\u4e00\u4e2a\u63d2\u4ef6\uff0c\u53d1\u73b0\u8c03\u5ea6\u7ed3\u679c\u4e0d\u7b26\u5408\u9884\u671f\uff0c\u4f46\u662f\u6709\u4e0d\u77e5\u9053\u54ea\u91cc\u51fa\u4e86\u95ee\u9898\u3002\u00a0\u00a0\u8981\u8bca\u65ad\u5206\u6790\u8fd9\u4e9b\u95ee\u9898\uff0c\u9664\u4e86\u8981\u638c\u63e1 Kubernetes \u57fa\u672c\u7684\u8c03\u5ea6\u673a\u5236\u548c\u8d44\u6e90\u5206\u914d\u673a\u5236\u5916\uff0c\u8fd8\u9700\u8981\u8c03\u5ea6\u5668\u81ea\u8eab\u7ed9\u4e88\u652f\u6301\u3002\u4f46\u662f Kubernetes kube-scheduler \u63d0\u4f9b\u7684\u8bca\u65ad\u80fd\u529b\u6bd4\u8f83\u6709\u9650\uff0c\u6709\u65f6\u5019\u751a\u81f3\u6ca1\u6709\u4ec0\u4e48\u65e5\u5fd7\u53ef\u4ee5\u67e5\u770b\u3002kube-scheduler \u539f\u751f\u662f\u652f\u6301\u901a\u8fc7 HTTP \u66f4\u6539\u65e5\u5fd7\u7b49\u7ea7\uff0c\u53ef\u4ee5\u83b7\u5f97\u66f4\u591a\u65e5\u5fd7\u4fe1\u606f\uff0c\u4f8b\u5982\u6267\u884c\u5982\u4e0b\u547d\u4ee4\u53ef\u4ee5\u66f4\u6539\u65e5\u5fd7\u7b49\u7ea7\u5230 5\uff1a\u00a0$ curl -X PUT schedulerLeaderIP:10251\/debug\/flags\/v &#8211;data &#8216;5&#8217; <br \/>successfully set klog.logging.verbosity to 5\u00a0Koordinator \u9488\u5bf9\u8fd9\u4e9b\u95ee\u9898\uff0c\u5b9e\u73b0\u4e86\u4e00\u5957 Restful API \uff0c\u5e2e\u52a9\u7528\u6237\u63d0\u5347\u95ee\u9898\u8bca\u65ad\u5206\u6790\u7684\u6548\u7387\u00a0\u5206\u6790 Score \u7ed3\u679c\u00a0PUT \/debug\/flags\/s \u00a0\u5141\u8bb8\u7528\u6237\u6253\u5f00 Debug Score \u5f00\u5173\uff0c\u5728\u6253\u5206\u7ed3\u675f\u540e\uff0c\u4ee5Markdown \u683c\u5f0f\u6253\u5370 TopN \u8282\u70b9\u5404\u4e2a\u63d2\u4ef6\u7684\u5206\u503c\u3002\u4f8b\u5982\uff1a\u00a0$ curl -X PUT schedulerLeaderIP:10251\/debug\/flags\/s &#8211;data &#8216;100&#8217;<\/p>\n<table>\n<thead>\n<tr>\n<th>successfully set debugTopNScores to 100\u00a0\u5f53\u6709\u65b0 Pod \u8c03\u5ea6\u65f6\uff0c\u89c2\u5bdf scheduler log \u53ef\u4ee5\u770b\u5230\u5982\u4e0b\u4fe1\u606f\u00a0<\/th>\n<th>#<\/th>\n<th>Pod<\/th>\n<th align=\"right\">Node<\/th>\n<th align=\"right\">Score<\/th>\n<th align=\"right\">ImageLocality<\/th>\n<th align=\"right\">InterPodAffinity<\/th>\n<th align=\"right\">LoadAwareScheduling<\/th>\n<th align=\"right\">NodeAffinity<\/th>\n<th align=\"right\">NodeNUMAResource<\/th>\n<th align=\"right\">NodeResourcesBalancedAllocation<\/th>\n<th align=\"right\">NodeResourcesFit<\/th>\n<th align=\"right\">PodTopologySpread<\/th>\n<th align=\"right\">Reservation<\/th>\n<th>TaintToleration<\/th>\n<th> <\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>0<\/td>\n<td>default\/curlimage-545745d8f8-rngp7<\/td>\n<td>cn-hangzhou.10.0.4.51<\/td>\n<td align=\"right\">577<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">87<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">96<\/td>\n<td align=\"right\">94<\/td>\n<td align=\"right\">200<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">100<\/td>\n<\/tr>\n<tr>\n<td>1<\/td>\n<td>default\/curlimage-545745d8f8-rngp7<\/td>\n<td>cn-hangzhou.10.0.4.50<\/td>\n<td align=\"right\">574<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">85<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">96<\/td>\n<td align=\"right\">93<\/td>\n<td align=\"right\">200<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">100<\/td>\n<\/tr>\n<tr>\n<td>2<\/td>\n<td>default\/curlimage-545745d8f8-rngp7<\/td>\n<td>cn-hangzhou.10.0.4.19<\/td>\n<td align=\"right\">541<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">55<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">95<\/td>\n<td align=\"right\">91<\/td>\n<td align=\"right\">200<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">100<\/td>\n<\/tr>\n<tr>\n<td>3<\/td>\n<td>default\/curlimage-545745d8f8-rngp7<\/td>\n<td>cn-hangzhou.10.0.4.18<\/td>\n<td align=\"right\">487<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">15<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">90<\/td>\n<td align=\"right\">82<\/td>\n<td align=\"right\">200<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">100<\/td>\n<td>\u00a0\u627e\u4e2a Markdown \u5de5\u5177\uff0c\u5c31\u53ef\u4ee5\u8f6c\u4e3a\u5982\u4e0b\u8868\u683c\uff1a\u00a0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042746719\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u8c03\u5ea6\u63d2\u4ef6\u5bfc\u51fa\u5185\u90e8\u72b6\u6001\u00a0\u50cf koord-scheduler \u5185\u90e8\u7684 NodeNUMAResource \u3001 DeviceShare \u548c ElasticQuota \u7b49\u63d2\u4ef6\u5185\u90e8\u90fd\u6709\u7ef4\u62a4\u4e00\u4e9b\u72b6\u6001\u5e2e\u52a9\u8c03\u5ea6\u3002koord-scheduler \u81ea\u5b9a\u4e49\u4e86\u4e00\u4e2a\u65b0\u7684\u63d2\u4ef6\u6269\u5c55\u63a5\u53e3\uff08\u5b9a\u4e49\u89c1\u4e0b\u6587\uff09\uff0c\u5e76\u4f1a\u5728\u521d\u59cb\u5316\u63d2\u4ef6\u540e\uff0c\u8bc6\u522b\u8be5\u63d2\u4ef6\u662f\u5426\u5b9e\u73b0\u4e86\u8be5\u63a5\u53e3\u5e76\u8c03\u7528\u8be5\u63a5\u53e3\uff0c\u8ba9\u63d2\u4ef6\u6ce8\u5165\u9700\u8981\u66b4\u9732\u7684 RestfulAPI\u3002\u4ee5 NodeNUMAResource \u63d2\u4ef6\u4e3a\u4f8b\uff0c\u4f1a\u63d0\u4f9b\/cpuTopologyOptions\/:nodeName \u548c\/availableCPUs\/:nodeName \u4e24\u4e2a Endpoints\uff0c\u53ef\u4ee5\u67e5\u770b\u63d2\u4ef6\u5185\u90e8\u8bb0\u5f55\u7684 CPU \u62d3\u6251\u4fe1\u606f\u548c\u5206\u914d\u7ed3\u679c\u3002\u00a0type APIServiceProvider interface {<br \/>  RegisterEndpoints(group *gin.RouterGroup)<br \/>}\u00a0\u7528\u6237\u5728\u4f7f\u7528\u65f6\uff0c\u6309\u7167 \/apis\/v1\/plugins\/&lt;pluginName&gt;\/&lt;pluginEndpoints&gt;\u65b9 \u5f0f\u6784\u5efa URL \u67e5\u770b\u6570\u636e\uff0c\u4f8b\u5982\u8981\u67e5\u770b \/cpuTopologyOptions\/:nodeName\uff1a\u00a0$ curl schedulerLeaderIP:10252\/apis\/v1\/plugins\/NodeNUMAResources\/cpuTopologyOptions\/node-1<br \/>{&#8220;cpuTopology&#8221;:{&#8220;numCPUs&#8221;:32,&#8221;numCores&#8221;:16,&#8221;numNodes&#8221;:1,&#8221;numSockets&#8221;:1,&#8221;cpuDetails&#8221;:&#8230;.\u00a0\u67e5\u770b\u5f53\u524d\u652f\u6301\u7684\u63d2\u4ef6 API\u00a0\u4e3a\u4e86\u65b9\u4fbf\u5927\u5bb6\u4f7f\u7528\uff0ckoord-scheduler \u63d0\u4f9b\u4e86 \/apis\/v1\/__services__ \u67e5\u770b\u652f\u6301\u7684 API Endpoints\u00a0$ curl schedulerLeaderIP:10251\/apis\/v1\/__services__<br \/>{<\/p>\n<pre><code>&quot;GET&quot;: [\n    &quot;\/apis\/v1\/__services__&quot;,\n    &quot;\/apis\/v1\/nodes\/:nodeName&quot;,\n    &quot;\/apis\/v1\/plugins\/Coscheduling\/gang\/:namespace\/:name&quot;,\n    &quot;\/apis\/v1\/plugins\/DeviceShare\/nodeDeviceSummaries&quot;,\n    &quot;\/apis\/v1\/plugins\/DeviceShare\/nodeDeviceSummaries\/:name&quot;,\n    &quot;\/apis\/v1\/plugins\/ElasticQuota\/quota\/:name&quot;,\n    &quot;\/apis\/v1\/plugins\/NodeNUMAResource\/availableCPUs\/:nodeName&quot;,\n    &quot;\/apis\/v1\/plugins\/NodeNUMAResource\/cpuTopologyOptions\/:nodeName&quot;\n]<\/code><\/pre>\n<p>}\u00a0\u66f4\u5b89\u5168\u7684\u91cd\u8c03\u5ea6\u00a0\u5728 Koordinator v0.6 \u7248\u672c\u4e2d\u6211\u4eec\u53d1\u5e03\u4e86\u5168\u65b0\u7684 koord-descheduler\uff0c\u652f\u6301\u63d2\u4ef6\u5316\u5b9e\u73b0\u9700\u8981\u7684\u91cd\u8c03\u5ea6\u7b56\u7565\u548c\u81ea\u5b9a\u4e49\u9a71\u9010\u673a\u5236\uff0c\u5e76\u5185\u7f6e\u4e86\u9762\u5411 PodMigrationJob \u7684\u8fc1\u79fb\u63a7\u5236\u5668\uff0c\u901a\u8fc7 Koordinator Reservation \u673a\u5236\u9884\u7559\u8d44\u6e90\uff0c\u786e\u4fdd\u6709\u8d44\u6e90\u7684\u60c5\u51b5\u4e0b\u53d1\u8d77\u9a71\u9010\u3002\u89e3\u51b3\u4e86 Pod \u88ab\u9a71\u9010\u540e\u65e0\u8d44\u6e90\u53ef\u7528\u5f71\u54cd\u5e94\u7528\u7684\u53ef\u7528\u6027\u95ee\u9898\u3002\u00a0Koordinator v0.7 \u7248\u672c\u4e2d\uff0ckoord-descheduler \u5b9e\u73b0\u4e86\u66f4\u5b89\u5168\u7684\u91cd\u8c03\u5ea6\u00a0\u652f\u6301 Evict \u9650\u6d41\uff0c\u7528\u6237\u53ef\u4ee5\u6839\u636e\u9700\u8981\u914d\u7f6e\u9650\u6d41\u7b56\u7565\uff0c\u4f8b\u5982\u5141\u8bb8\u6bcf\u5206\u949f\u9a71\u9010\u591a\u5c11\u4e2a Pod\u652f\u6301\u914d\u7f6e Namespace \u7070\u5ea6\u91cd\u8c03\u5ea6\u80fd\u529b\uff0c\u8ba9\u7528\u6237\u53ef\u4ee5\u66f4\u653e\u5fc3\u7684\u7070\u5ea6\u652f\u6301\u6309\u7167 Node\/Namespace \u914d\u7f6e\u9a71\u9010\u6570\u91cf\uff0c\u4f8b\u5982\u914d\u7f6e\u8282\u70b9\u7ef4\u5ea6\u6700\u591a\u53ea\u9a71\u9010\u4e24\u4e2a\uff0c\u90a3\u4e48\u5373\u4f7f\u6709\u63d2\u4ef6\u8981\u6c42\u9a71\u9010\u8be5\u8282\u70b9\u4e0a\u7684\u66f4\u591a Pod\uff0c\u4f1a\u88ab\u62d2\u7edd\u3002\u611f\u77e5 Workload \uff0c\u5982\u679c\u4e00\u4e2a Workload \u6b63\u5728\u53d1\u5e03\u3001\u7f29\u5bb9\u3001\u5df2\u7ecf\u6709\u4e00\u5b9a\u91cf\u7684 Pod \u6b63\u5728\u88ab\u9a71\u9010\u6216\u8005\u4e00\u4e9b Pod NotReady\uff0c\u91cd\u8c03\u5ea6\u5668\u4f1a\u62d2\u7edd\u65b0\u7684\u91cd\u8c03\u5ea6\u8bf7\u6c42\u3002\u76ee\u524d\u652f\u6301\u539f\u751f\u7684 Deployment\uff0cStatefulSet \u4ee5\u53ca Kruise \u00a0CloneSet\uff0cKruise AdvancedStatefulSet\u3002\u00a0\u00a0\u540e\u7eed\u91cd\u8c03\u5ea6\u5668\u8fd8\u4f1a\u63d0\u5347\u516c\u5e73\u6027\uff0c\u9632\u6b62\u4e00\u76f4\u91cd\u590d\u7684\u91cd\u8c03\u5ea6\u540c\u4e00\u4e2a workload \uff0c\u5c3d\u91cf\u964d\u4f4e\u91cd\u8c03\u5ea6\u5bf9\u5e94\u7528\u7684\u53ef\u7528\u6027\u7684\u5f71\u54cd\u3002\u00a0\u5176\u4ed6\u6539\u52a8\u00a0Koordinator \u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86 CPU \u7cbe\u7ec6\u5316\u8c03\u5ea6\u80fd\u529b\uff0c\u5b8c\u5168\u517c\u5bb9 kubelet ( &lt;= v1.22) CPU Manager static \u7b56\u7565\u3002\u8c03\u5ea6\u5668\u5206\u914d CPU \u65f6\u4f1a\u907f\u514d\u5206\u914d\u88ab kubelet \u9884\u7559\u7684 CPU\uff0c\u5355\u673a\u4fa7 koordlet \u5b8c\u6574\u9002\u914d\u4e86 kubelet \u4ece 1.18 \u5230 1.22 \u7248\u672c\u7684\u5206\u914d\u7b56\u7565\uff0c\u6709\u6548\u907f\u514d\u4e86 CPU \u51b2\u7a81\u3002\u8d44\u6e90\u9884\u7559\u673a\u5236\u652f\u6301 AllocateOnce \u8bed\u4e49\uff0c\u6ee1\u8db3\u5355\u6b21\u9884\u7559\u573a\u666f\u3002\u5e76\u6539\u8fdb\u4e86 Reservation \u72b6\u6001\u8bed\u4e49\uff0c\u66f4\u52a0\u51c6\u786e\u63cf\u8ff0 Reservation \u5bf9\u8c61\u5f53\u524d\u7684\u72b6\u6001\u3002\u6539\u8fdb\u4e86\u79bb\u7ebf\u8d44\u6e90(Batch CPU\/Memory) \u7684\u58f0\u660e\u65b9\u5f0f\uff0c\u652f\u6301 limit \u5927\u4e8e request \u7684\u8d44\u6e90\u63cf\u8ff0\u5f62\u5f0f\uff0c\u53ef\u4ee5\u65b9\u4fbf\u539f burstable \u7c7b\u578b\u7684\u4efb\u52a1\u76f4\u63a5\u8f6c\u6362\u4e3a\u6df7\u90e8\u6a21\u5f0f\u8fd0\u884c\u3002\u00a0\u00a0\u4f60\u53ef\u4ee5\u901a\u8fc7\u00a0Github release[3]\u9875\u9762\uff0c\u6765\u67e5\u770b\u66f4\u591a\u7684\u6539\u52a8\u4ee5\u53ca\u5b83\u4eec\u7684\u4f5c\u8005\u4e0e\u63d0\u4ea4\u8bb0\u5f55\u3002\u00a0\u793e\u533a\u53c2\u4e0e\u00a0\u975e\u5e38\u6b22\u8fce\u4f60\u901a\u8fc7 Github\/Slack\/\u9489\u9489\/\u5fae\u4fe1 \u7b49\u65b9\u5f0f\u52a0\u5165\u6211\u4eec\u6765\u53c2\u4e0e Koordinator \u5f00\u6e90\u793e\u533a\u3002\u4f60\u662f\u5426\u5df2\u7ecf\u6709\u4e00\u4e9b\u5e0c\u671b\u4e0e\u6211\u4eec\u793e\u533a\u4ea4\u6d41\u7684\u5185\u5bb9\u5462\uff1f\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u6e20\u9053\u53c2\u4e0e\u8ba8\u8bba\uff1a\u00a0\u52a0\u5165\u793e\u533a\u00a0Slack channel[4]\u00a0(English)\u52a0\u5165\u793e\u533a\u9489\u9489\u7fa4\uff1a\u641c\u7d22\u7fa4\u53f7 33383887 (Chinese) \u6216\u8005\u626b\u63cf\u4e0b\u65b9\u4e8c\u7ef4\u7801\u00a0\u00a0<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042746720\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u76f8\u5173\u94fe\u63a5\u00a0[1] Koordinator<a href=\"https:\/\/link.segmentfault.com\/?enc=rBvZ4FYkjpIpsBNtJMol5w%3D%3D.qIrhYjSbMBV8rCGot3Nab8JarXPOk6TnVcZPM%2FJLMhM%3D\" rel=\"nofollow\">https:\/\/koordinator.sh\/<\/a>\u00a0[2] Koordinator 0.6 Release Note<a href=\"https:\/\/link.segmentfault.com\/?enc=EwH7zvqIYGZvAQ65gdjn9Q%3D%3D.2n4rhu%2B6oMfYTzxLuOG605q2U6nrZOA14HhPafpz7jgUjbeeoa%2BWwBU50Ue9GNRuQ5Scqf6J3OWo%2F9AiHt%2Fxzg%3D%3D\" rel=\"nofollow\">https:\/\/mp.weixin.qq.com\/s\/Yd&#8230;<\/a>\u00a0[3]\u00a0Github Release<a href=\"https:\/\/link.segmentfault.com\/?enc=u9tgkymy3ZK86M9qC8pILQ%3D%3D.mU5fvjXxt2XxgWlly48P2%2BxMNrMNeBH48snhVwdibmjCK15NzqmeIcMZgN570RV2otpgLkgYUVxRDHzMH4hrGJwXgn4UYKtF8EkwRkuyEcQ%3D\" rel=\"nofollow\">https:\/\/github.com\/koordinato&#8230;<\/a>\u00a0[4]\u00a0Slack Channel<a href=\"https:\/\/link.segmentfault.com\/?enc=VEKAmxY3oKIOqciqIa9Uiw%3D%3D.A4QXcB6sPCr0hmKPOiYC5yTxLLFhHDJReTIA%2B0km%2BxlLi9qrg6uP8i%2FF3QwopmiHpOaRZoXpXcHYACM1KxLeq%2By%2BNOc1F%2FJxrd7g3sExVSXtW9bZe1MR%2FREeXNBhr7fz\" rel=\"nofollow\">https:\/\/join.slack.com\/t\/koor&#8230;<\/a>\u00a0[5] Design: Gang Scheduling<a href=\"https:\/\/link.segmentfault.com\/?enc=gnmRHQArcZA39s2hdLXjAA%3D%3D.ileZMOzCgybRvFhVaicqwStfC7BTsu4ajXf2OfZ6mf40%2Fgtr%2Bi6ILfMm%2FD6RWWcTSyqYA4Px6hOyXXfX7J9o9LTxXYzxdif4jf9rqD0%2FJxPv0S2hGnQ9l5ZjPaT8i9AelfCBhM%2B0UuiLOW%2B5k9f7CA%3D%3D\" rel=\"nofollow\">https:\/\/github.com\/koordinato&#8230;<\/a>\u00a0[6] Design: Multi Hierarchy ElasticQuota Management<a href=\"https:\/\/link.segmentfault.com\/?enc=AXwp520VGqBIDLtSzmucuA%3D%3D.%2FoQZLt7Q%2BgyyK798GObw9gxMGGmFYW9rmjeQu2pSHQ6Du8j4eRb%2B%2F0f%2FFo2hNyx%2FmREB7Aph1XQdhO%2BmlEe1Bq2RqzQiqFd32p%2BdoQGUmE9x1CnLJ9%2BThNl6WsESkwwv5uUDRxi7YtA2eNDsMTcZWPybSzlcPc5WRz%2BQuw%2FTPvYCvHv06VbBqCYxZg9oyTGT\" rel=\"nofollow\">https:\/\/github.com\/koordinato&#8230;<\/a>\u00a0[7] Design: Fine-grained Device Scheduling<a href=\"https:\/\/link.segmentfault.com\/?enc=Zg81KkHxlPUq49zitwiVbQ%3D%3D.owMbzk%2FUAiOUuaPttVl6j6DIx%2FtjOPLYOVHpHWZM9L5wLMAJgDtC9xFEFfTva10sjLYZuQht6D8PNqw2fHPVzyOQhgEN%2F8Hov8685IO9X768WU2jtccfHSGgHwkviMRfNjCoJzFaxC02%2B8MuPekaqbB5mRhQw%2FDlCLIWvUglYa4%3D\" rel=\"nofollow\">https:\/\/github.com\/koordinato&#8230;<\/a>\u00a0[8] \u4e91\u539f\u751f\u6df7\u90e8\u7cfb\u7edf Koordinator \u67b6\u6784\u8be6\u89e3<a href=\"https:\/\/link.segmentfault.com\/?enc=Q1X%2B9xUCOSkafNOoQtTPrQ%3D%3D.p1CFzWI2KLD9w1PRNNIDymZ22wunWbWjFX%2FEwKUyV1CAN2tL%2FvDXD%2Fe3tC60VRv9%2BGgMaiO0tyz9YHderKcVGg%3D%3D\" rel=\"nofollow\">https:\/\/mp.weixin.qq.com\/s\/y8&#8230;<\/a>\u00a0\u70b9\u51fb\u6b64\u5904\uff0c\u7acb\u5373\u4e86\u89e3 Koordinator \u9879\u76ee\uff01\u539f\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=L3hBa6%2BOvwer6%2FNMj%2FwZdA%3D%3D.u3RqfInSI64cW2j%2BxWDn2ETjX0YztjxsNpKB7GEqygw8Q9IWpDjgbYi4Avy99sjR\" rel=\"nofollow\">https:\/\/click.aliyun.com\/m\/10&#8230;<\/a>\u672c\u6587\u4e3a\u963f\u91cc\u4e91\u539f\u521b\u5185\u5bb9\uff0c\u672a\u7ecf\u5141\u8bb8\u4e0d\u5f97\u8f6c\u8f7d\u3002<\/p>\n<p><a href=\"https:\/\/segmentfault.com\/a\/1190000042746713\">https:\/\/segmentfault.com\/a\/1190000042746713<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7b80\u4ecb\uff1a\u00a0\u5728\u8fd9\u4e2a\u7248\u672c\u4e2d\u7740\u91cd\u5efa\u8bbe\u4e86\u673a\u5668\u5b66\u4e60\u3001\u5927\u6570\u636e\u573a\u666f\u9700\u8981\u7684\u4efb\u52a1\u8c03\u5ea6\u80fd\u529b\uff0c\u4f8b\u5982 Coscheduling\u3001Elas [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10006812","post","type-post","status-publish","format-standard","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006812","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10006812"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006812\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10006812"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10006812"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10006812"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}