Yarn公平调度器Fair Scheduler根据用户组分配资源池

假设在生产环境Yarn中，总共有四类用户需要使用集群，开发用户、测试用户、业务1用户、业务2用户。为了使其提交的任务不受影响，我们在Yarn上规划配置了五个资源池，分别为 dev_group（开发用户组资源池）、test_group（测试用户组资源池）、business1_group（业务1用户组资源池）、business2_group（业务2用户组资源池）、default（只分配了极少资源）。并根据实际业务情况，为每个资源池分配了相应的资源及优先级等。

更多精彩内容点我学

ResourceManager上fair-scheduler.xml配置如下：

1. <?xml version=“1.0”?>
2. <allocations>  
3.   <!– users max running apps –>
4.   <userMaxAppsDefault>30</userMaxAppsDefault>
5. <queue name=“root”>
6. <aclSubmitApps> </aclSubmitApps>
7. <aclAdministerApps> </aclAdministerApps>
8.  
9.   <queue name=“default”>
10.           <minResources>2000mb,1vcores</minResources>
11.           <maxResources>10000mb,1vcores</maxResources>
12.           <maxRunningApps>1</maxRunningApps>
13.           <schedulingMode>fair</schedulingMode>
14.           <weight>0.5</weight>
15.           <aclSubmitApps>*</aclSubmitApps>
16.   </queue>
17.        
18.   <queue name=“dev_group”>
19.           <minResources>200000mb,33vcores</minResources>
20.           <maxResources>300000mb,90vcores</maxResources>
21.           <maxRunningApps>150</maxRunningApps>
22.           <schedulingMode>fair</schedulingMode>
23.           <weight>2.5</weight>
24.           <aclSubmitApps> dev_group</aclSubmitApps>
25.           <aclAdministerApps> hadoop,dev_group</aclAdministerApps>
26.   </queue>
27.                                                                                                                                                   
28.   <queue name=“test_group”>
29.           <minResources>70000mb,20vcores</minResources>
30.           <maxResources>95000mb,25vcores</maxResources>
31.           <maxRunningApps>60</maxRunningApps>
32.           <schedulingMode>fair</schedulingMode>
33.           <weight>1</weight>
34.           <aclSubmitApps> test_group</aclSubmitApps>
35.           <aclAdministerApps> hadoop,test_group</aclAdministerApps>
36.   </queue>
37.                                                                           
38.   <queue name=“business1_group”>
39.           <minResources>75000mb,15vcores</minResources>
40.           <maxResources>100000mb,20vcores</maxResources>
41.           <maxRunningApps>80</maxRunningApps>
42.           <schedulingMode>fair</schedulingMode>
43.           <weight>1</weight>
44.           <aclSubmitApps> business1_group</aclSubmitApps>
45.           <aclAdministerApps> hadoop,business1_group</aclAdministerApps>
46.   </queue>
47.                                                              
48.                                                                           
49.   <queue name=“business2_group”>
50.       <minResources>75000mb,15vcores</minResources>
51.       <maxResources>102400mb,20vcores</maxResources>
52.       <maxRunningApps>80</maxRunningApps>
53.       <schedulingMode>fair</schedulingMode>
54.       <weight>1</weight>
55.       <aclSubmitApps> business2_group</aclSubmitApps>
56.       <aclAdministerApps> hadoop,business2_group</aclAdministerApps>
57.   </queue>
58.  
59. </queue>
60.   <queuePlacementPolicy>
61.       <rule name=“primaryGroup” create=“false” />
62.       <rule name=“secondaryGroupExistingQueue” create=“false” />
63.       <rule name=“default” />
64.   </queuePlacementPolicy>
65.  
66. </allocations>更多精彩内容点我学​​​​​​​

这样，每个用户组下的用户提交任务时候，会到相应的资源池中，而不影响其他业务。

需要注意的是，所有客户端提交任务的用户和用户组的对应关系，需要维护在ResourceManager上，ResourceManager在分配资源池时候，是从ResourceManager上读取用户和用户组的对应关系的，否则就会被分配到default资源池。在日志中出现”UserGroupInformation: No groups available for user”类似的警告。而客户端机器上的用户对应的用户组无关紧要。

每次在ResourceManager上新增用户或者调整资源池配额后，需要执行下面的命令刷新使其生效：

1. yarn rmadmin –refreshQueues
2. yarn rmadmin –refreshUserToGroupsMappings