Configuring the Fair Scheduler

The Fair Scheduler is the Cloudera recommended scheduler option. The Fair Scheduler controls how resources are allocated to pools (or queues) and how jobs are assigned to pools. Jobs can also be explicitly submitted to pools; to submit an job to a specific pool, you specify the mapreduce.job.queuename property.

Pools have policies for preempting running jobs, so that when there is contention for resources, jobs that have high priority or have been waiting a long time to run are allowed to run.

Fair Scheduler configuration is maintained in two files: yarn-site.xml and fair-scheduler.xml. Detailed information on the available properties is available at Fair Scheduler Configuration. When you change the contents of yarn-site.xml, you must restart the YARN service.

In Cloudera Manager the Dynamic Resource Pools Configuration screen provides an enhanced interface for configuring the Fair Scheduler. In addition to allowing you to configure resource allocation properties, you can define schedules for changing the values of the properties. Cloudera Manager automatically updates Fair Scheduler configuration files according to the schedule.
yarn-site.xml Properties
Property Description
yarn.scheduler.fair.allow-undeclared-pools When set to true, the Fair Scheduler uses the username as the default pool name, in the event that a pool name is not specified. When set to false, all applications are run in a shared pool, called default.

Default: true.
yarn.scheduler.fair.user-as-default-queue When set to true, pools specified in applications but not explicitly configured, are created at runtime with default settings. When set to false, applications specifying pools not explicitly configured run in a pool named default. This setting applies when an application explicitly specifies a pool and when the application runs in a pool named with the username associated with the application.

Default: true.
yarn.scheduler.fair.preemption When enabled, under certain conditions, the Fair Scheduler preempts applications in other pools. Preemption guarantees that production applications are not starved while also allowing the cluster to be used for experimental and research applications. To minimize wasted computation, the Fair Scheduler preempts the most recently launched applications.

Default: false.
yarn.scheduler.fair.preemption.cluster-utilization-threshold The cluster utilization threshold above which preemption is triggered. If the cluster utilization is under this threshold, preemption is not triggered even if there are starved queues. The utilization is computed as the maximum ratio of usage to capacity among all resources.

Default: .8.
For example:
yarn-site.xml Example 
...
  <property>
    <name>yarn.scheduler.fair.allow-undeclared-pools</name>
    <value>true</value>
  </property>
  <property>
    <name>yarn.scheduler.fair.user-as-default-queue</name>
    <value>true</value>
  </property>
  <property>
    <name>yarn.scheduler.fair.preemption</name>
    <value>true</value>
  </property>
  <property>
    <name>yarn.scheduler.fair.preemption.cluster-utilization-threshold</name>
    <value>0.8</value>
  </property>
...
fair-scheduler.xml Properties
Element Subelement Description
queuePlacementPolicy   Policy for assigning jobs to resource pools. In Cloudera Manager this property is configured using placement rules.
userMaxAppsDefault   Default running app limit for a user whose limit is not otherwise specified. In Cloudera Manager this property is configured using user limits.
queueMaxAppsDefault   Default running app limit for pools; overridden by the equivalent element in a pool.
queueMaxAMShareDefault   Default ApplicationMaster resource limit for the pool; overridden by the equivalent element in a pool.
defaultFairSharePreemptionThreshold   Fair share preemption threshold for pools; overridden by the equivalent element in a pool. The threshold value is between 0 and 1. If set to x and the fair share of the pool is F, resources are preempted from other pools if the allocation is less than (x * F).
defaultFairSharePreemptionTimeout   Default number of seconds a resource pool is under its fair share before it will preempt containers to take resources from other resource pools; overridden by the equivalent element in a pool. If this parameter is not set, and fairSharePreemptionTimeout is not set for a given queue or its ancestor queues, pre-emption by this queue will never occur, even if pre-emption is enabled. Default timeout is 2^64 milliseconds.
defaultMinSharePreemptionTimeout   Default number of seconds a resource pool is under its minimum share before it will preempt containers to take resources from other resource pools; overridden by the equivalent element in a pool.
defaultQueueSchedulingPolicy   Default scheduling policy for pools; overridden by the equivalent element in a pool.

Default: drf.
queue   Name of a dynamic resource pool.
  weight Weight given to the resource pool when determining how to allocate resources relative to other resource pools. In Cloudera Manager this property is configured using configuration sets.
  schedulingPolicy Policy to determine how resources are allocated to the resource pool: fair, fifo, or drf.
  aclSubmitApps Users and groups that can submit jobs to the pool.
  aclAdministerApps Users and groups that can administer the pool.
  minResources, maxResources Minimum and maximum share of resources that can allocated to the resource pool in the form X mb, Y vcores. Values computed by the weight settings are limited by (or constrained by) the minimum and maximum values.
  maxAMShare Fraction of the resource pool's fair share that can be used to run ApplicationMasters. For example, if set to 1.0, then ApplicationMasters in the pool can take up to 100% of both the memory and CPU fair share. The value of -1.0 disables this feature, and the ApplicationMaster share is not checked. The default value is 0.5.
  maxRunningApps See default elements.
  fairSharePreemptionThreshold See default elements.
  fairSharePreemptionTimeout See default elements.
  minSharePreemptionTimeout See default elements.
For example:
fair-scheduler.xml 
<allocations>
    <queue name="root">
        <weight>1.0</weight>
        <schedulingPolicy>drf</schedulingPolicy>
        <aclSubmitApps> </aclSubmitApps>
        <aclAdministerApps>*</aclAdministerApps>
        <queue name="production">
            <minResources>1024 mb, 10 vcores</minResources>
            <maxResources>5120 mb, 20 vcores</maxResources>
            <weight>4.0</weight>
            <schedulingPolicy>drf</schedulingPolicy>
            <aclSubmitApps>*</aclSubmitApps>
            <aclAdministerApps>*</aclAdministerApps>
        </queue>
        <queue name="development">
            <weight>1.0</weight>
            <schedulingPolicy>drf</schedulingPolicy>
            <aclSubmitApps>*</aclSubmitApps>
            <aclAdministerApps>*</aclAdministerApps>
        </queue>
    </queue>
    <defaultQueueSchedulingPolicy>drf</defaultQueueSchedulingPolicy>
    <queuePlacementPolicy>
        <rule name="specified" create="true"/>
        <rule name="user" create="true"/>
    </queuePlacementPolicy>
</allocations>

Dynamic resource pools allow you to configure scheduler properties. See Configuring Default YARN Fair Scheduler Properties.