Hadoop MapReduce Next Generation - Fair Scheduler
This document describes the FairScheduler, a pluggable scheduler for Hadoop that allows YARN applications to share resources in large clusters fairly.
Fair scheduling is a method of assigning resources to applications such that all apps get, on average, an equal share of resources over time. Hadoop NextGen is capable of scheduling multiple resource types. By default, the Fair Scheduler bases scheduling fairness decisions only on memory. It can be configured to schedule with both memory and CPU, using the notion of Dominant Resource Fairness developed by Ghodsi et al. When there is a single app running, that app uses the entire cluster. When other apps are submitted, resources that free up are assigned to the new apps, so that each app eventually on gets roughly the same amount of resources. Unlike the default Hadoop scheduler, which forms a queue of apps, this lets short apps finish in reasonable time while not starving long-lived apps. It is also a reasonable way to share a cluster between a number of users. Finally, fair sharing can also work with app priorities - the priorities are used as weights to determine the fraction of total resources that each app should get.
The scheduler organizes apps further into "queues", and shares resources fairly between these queues. By default, all users share a single queue, named "default". If an app specifically lists a queue in a container resource request, the request is submitted to that queue. It is also possible to assign queues based on the user name included with the request through configuration. Within each queue, a scheduling policy is used to share resources between the running apps. The default is memory-based fair sharing, but FIFO and multi-resource with Dominant Resource Fairness can also be configured. Queues can be arranged in a hierarchy to divide resources and configured with weights to share the cluster in specific proportions.
In addition to providing fair sharing, the Fair Scheduler allows assigning guaranteed minimum shares to queues, which is useful for ensuring that certain users, groups or production applications always get sufficient resources. When a queue contains apps, it gets at least its minimum share, but when the queue does not need its full guaranteed share, the excess is split between other running apps. This lets the scheduler guarantee capacity for queues while utilizing resources efficiently when these queues don't contain applications.
The Fair Scheduler lets all apps run by default, but it is also possible to limit the number of running apps per user and per queue through the config file. This can be useful when a user must submit hundreds of apps at once, or in general to improve performance if running too many apps at once would cause too much intermediate data to be created or too much context-switching. Limiting the apps does not cause any subsequently submitted apps to fail, only to wait in the scheduler's queue until some of the user's earlier apps finish.
The fair scheduler supports hierarchical queues. All queues descend from a queue named "root". Available resources are distributed among the children of the root queue in the typical fair scheduling fashion. Then, the children distribute the resources assigned to them to their children in the same fashion. Applications may only be scheduled on leaf queues. Queues can be specified as children of other queues by placing them as sub-elements of their parents in the fair scheduler allocation file.
A queue's name starts with the names of its parents, with periods as separators. So a queue named "queue1" under the root queue, would be referred to as "root.queue1", and a queue named "queue2" under a queue named "parent1" would be referred to as "root.parent1.queue2". When referring to queues, the root part of the name is optional, so queue1 could be referred to as just "queue1", and a queue2 could be referred to as just "parent1.queue2".
Additionally, the fair scheduler allows setting a different custom policy for each queue to allow sharing the queue's resources in any which way the user wants. A custom policy can be built by extending org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.SchedulingPolicy. FifoPolicy, FairSharePolicy (default), and DominantResourceFairnessPolicy are built-in and can be readily used.
Certain add-ons are not yet supported which existed in the original (MR1) Fair Scheduler. Among them, is the use of a custom policies governing priority "boosting" over certain apps.
The Fair Scheduler allows administrators to configure policies that automatically place submitted applications into appropriate queues. Placement can depend on the user and groups of the submitter and the requested queue passed by the application. A policy consists of a set of rules that are applied sequentially to classify an incoming application. Each rule either places the app into a queue, rejects it, or continues on to the next rule. Refer to the allocation file format below for how to configure these policies.
To use the Fair Scheduler first assign the appropriate scheduler class in yarn-site.xml:
Customizing the Fair Scheduler typically involves altering two files. First, scheduler-wide options can be set by adding configuration properties in the yarn-site.xml file in your existing configuration directory. Second, in most cases users will want to create an allocation file listing which queues exist and their respective weights and capacities. The allocation file is reloaded every 10 seconds, allowing changes to be made on the fly.
Properties that can be placed in yarn-site.xml
- Path to allocation file. An allocation file is an XML manifest describing queues and their properties, in addition to certain policy defaults. This file must be in the XML format described in the next section. If a relative path is given, the file is searched for on the classpath (which typically includes the Hadoop conf directory). Defaults to fair-scheduler.xml.
- Whether to use the username associated with the allocation as the default queue name, in the event that a queue name is not specified. If this is set to "false" or unset, all jobs have a shared default queue, named "default". Defaults to true. If a queue placement policy is given in the allocations file, this property is ignored.
- Whether to use preemption. Defaults to false.
- The utilization threshold after which preemption kicks in. The utilization is computed as the maximum ratio of usage to capacity among all resources. Defaults to 0.8f.
- Whether to assign shares to individual apps based on their size, rather than providing an equal share to all apps regardless of size. When set to true, apps are weighted by the natural logarithm of one plus the app's total requested memory, divided by the natural logarithm of 2. Defaults to false.
- Whether to allow multiple container assignments in one heartbeat. Defaults to false.
- If assignmultiple is true, the maximum amount of containers that can be assigned in one heartbeat. Defaults to -1, which sets no limit.
- For applications that request containers on particular nodes, the number of scheduling opportunities since the last container assignment to wait before accepting a placement on another node. Expressed as a float between 0 and 1, which, as a fraction of the cluster size, is the number of scheduling opportunities to pass up. The default value of -1.0 means don't pass up any scheduling opportunities.
- For applications that request containers on particular racks, the number of scheduling opportunities since the last container assignment to wait before accepting a placement on another rack. Expressed as a float between 0 and 1, which, as a fraction of the cluster size, is the number of scheduling opportunities to pass up. The default value of -1.0 means don't pass up any scheduling opportunities.
- If this is true, new queues can be created at application submission time, whether because they are specified as the application's queue by the submitter or because they are placed there by the user-as-default-queue property. If this is false, any time an app would be placed in a queue that is not specified in the allocations file, it is placed in the "default" queue instead. Defaults to true. If a queue placement policy is given in the allocations file, this property is ignored.
- The interval at which to lock the scheduler and recalculate fair shares, recalculate demand, and check whether anything is due for preemption. Defaults to 500 ms.
Allocation file format
The allocation file must be in XML format. The format contains five types of elements:
- Queue elements, which represent queues. Queue elements can take an optional attribute 'type', which when set to 'parent' makes it a parent queue. This is useful when we want to create a parent queue without configuring any leaf queues. Each queue element may contain the following properties:
- minResources: minimum resources the queue is entitled to, in the form "X mb, Y vcores". For the single-resource fairness policy, the vcores value is ignored. If a queue's minimum share is not satisfied, it will be offered available resources before any other queue under the same parent. Under the single-resource fairness policy, a queue is considered unsatisfied if its memory usage is below its minimum memory share. Under dominant resource fairness, a queue is considered unsatisfied if its usage for its dominant resource with respect to the cluster capacity is below its minimum share for that resource. If multiple queues are unsatisfied in this situation, resources go to the queue with the smallest ratio between relevant resource usage and minimum. Note that it is possible that a queue that is below its minimum may not immediately get up to its minimum when it submits an application, because already-running jobs may be using those resources.
- maxResources: maximum resources a queue is allowed, in the form "X mb, Y vcores". For the single-resource fairness policy, the vcores value is ignored. A queue will never be assigned a container that would put its aggregate usage over this limit.
- maxRunningApps: limit the number of apps from the queue to run at once
- maxAMShare: limit the fraction of the queue's fair share that can be used to run application masters. This property can only be used for leaf queues. For example, if set to 1.0f, then AMs in the leaf queue can take up to 100% of both the memory and CPU fair share. The value of -1.0f will disable this feature and the amShare will not be checked. The default value is 0.5f.
- weight: to share the cluster non-proportionally with other queues. Weights default to 1, and a queue with weight 2 should receive approximately twice as many resources as a queue with the default weight.
- schedulingPolicy: to set the scheduling policy of any queue. The allowed values are "fifo"/"fair"/"drf" or any class that extends org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.SchedulingPolicy. Defaults to "fair". If "fifo", apps with earlier submit times are given preference for containers, but apps submitted later may run concurrently if there is leftover space on the cluster after satisfying the earlier app's requests.
- aclSubmitApps: a list of users and/or groups that can submit apps to the queue. Refer to the ACLs section below for more info on the format of this list and how queue ACLs work.
- aclAdministerApps: a list of users and/or groups that can administer a queue. Currently the only administrative action is killing an application. Refer to the ACLs section below for more info on the format of this list and how queue ACLs work.
- minSharePreemptionTimeout: number of seconds the queue is under its minimum share before it will try to preempt containers to take resources from other queues. If not set, the queue will inherit the value from its parent queue.
- fairSharePreemptionTimeout: number of seconds the queue is under its fair share threshold before it will try to preempt containers to take resources from other queues. If not set, the queue will inherit the value from its parent queue.
- fairSharePreemptionThreshold: the fair share preemption threshold for the queue. If the queue waits fairSharePreemptionTimeout without receiving fairSharePreemptionThreshold*fairShare resources, it is allowed to preempt containers to take resources from other queues. If not set, the queue will inherit the value from its parent queue.
- User elements, which represent settings governing the behavior of individual users. They can contain a single property: maxRunningApps, a limit on the number of running apps for a particular user.
- A userMaxAppsDefault element, which sets the default running app limit for any users whose limit is not otherwise specified.
- A defaultFairSharePreemptionTimeout element, which sets the fair share preemption timeout for the root queue; overridden by fairSharePreemptionTimeout element in root queue.
- A defaultMinSharePreemptionTimeout element, which sets the min share preemption timeout for the root queue; overridden by minSharePreemptionTimeout element in root queue.
- A defaultFairSharePreemptionThreshold element, which sets the fair share preemption threshold for the root queue; overridden by fairSharePreemptionThreshold element in root queue.
- A queueMaxAppsDefault element, which sets the default running app limit for queues; overriden by maxRunningApps element in each queue.
- A queueMaxAMShareDefault element, which sets the default AM resource limit for queue; overriden by maxAMShare element in each queue.
- A defaultQueueSchedulingPolicy element, which sets the default scheduling policy for queues; overriden by the schedulingPolicy element in each queue if specified. Defaults to "fair".
- A queuePlacementPolicy element, which contains a list of rule elements that tell the scheduler how to place incoming apps into queues. Rules are applied in the order that they are listed. Rules may take arguments. All rules accept the "create" argument, which indicates whether the rule can create a new queue. "Create" defaults to true; if set to false and the rule would place the app in a queue that is not configured in the allocations file, we continue on to the next rule. The last rule must be one that can never issue a continue. Valid rules are:
- specified: the app is placed into the queue it requested. If the app requested no queue, i.e. it specified "default", we continue.
- user: the app is placed into a queue with the name of the user who submitted it.
- primaryGroup: the app is placed into a queue with the name of the primary group of the user who submitted it.
- secondaryGroupExistingQueue: the app is placed into a queue with a name that matches a secondary group of the user who submitted it. The first secondary group that matches a configured queue will be selected.
- nestedUserQueue : the app is placed into a queue with the name of the user under the queue suggested by the nested rule. This is similar to ‘user’ rule,the difference being in 'nestedUserQueue' rule,user queues can be created under any parent queue, while 'user' rule creates user queues only under root queue. Note that nestedUserQueue rule would be applied only if the nested rule returns a parent queue.One can configure a parent queue either by setting 'type' attribute of queue to 'parent' or by configuring at least one leaf under that queue which makes it a parent. See example allocation for a sample use case.
- default: the app is placed into the queue specified in the 'queue' attribute of the default rule. If 'queue' attribute is not specified, the app is placed into 'root.default' queue.
- reject: the app is rejected.
An example allocation file is given here:
<!—- Queue 'secondary_group_queueue' is a parent queue and may have
user queues under it -->
<queue name="secondary_group_queue" type="parent">
<rule name="specified" />
<rule name="primaryGroup" create="false" />
<rule name="secondaryGroupExistingQueue" create="false" />
<rule name="default" queue="sample_queue"/>
Note that for backwards compatibility with the original FairScheduler, "queue" elements can instead be named as "pool" elements.
Queue Access Control Lists (ACLs)
Queue Access Control Lists (ACLs) allow administrators to control who may take actions on particular queues. They are configured with the aclSubmitApps and aclAdministerApps properties, which can be set per queue. Currently the only supported administrative action is killing an application. Anybody who may administer a queue may also submit applications to it. These properties take values in a format like "user1,user2 group1,group2" or " group1,group2". An action on a queue will be permitted if its user or group is in the ACL of that queue or in the ACL of any of that queue's ancestors. So if queue2 is inside queue1, and user1 is in queue1's ACL, and user2 is in queue2's ACL, then both users may submit to queue2.
Note: The delimiter is a space character. To specify only ACL groups, begin the value with a space character.
The root queue's ACLs are "*" by default which, because ACLs are passed down, means that everybody may submit to and kill applications from every queue. To start restricting access, change the root queue's ACLs to something other than "*".
The fair scheduler provides support for administration at runtime through a few mechanisms:
Modifying configuration at runtime
It is possible to modify minimum shares, limits, weights, preemption timeouts and queue scheduling policies at runtime by editing the allocation file. The scheduler will reload this file 10-15 seconds after it sees that it was modified.
Monitoring through web UI
Current applications, queues, and fair shares can be examined through the ResourceManager's web interface, at http://ResourceManager URL/cluster/scheduler.
The following fields can be seen for each queue on the web interface:
Moving applications between queues
The Fair Scheduler supports moving a running application to a different queue. This can be useful for moving an important application to a higher priority queue, or for moving an unimportant application to a lower priority queue. Apps can be moved by running "yarn application -movetoqueue appID -queue targetQueueName".
When an application is moved to a queue, its existing allocations become counted with the new queue's allocations instead of the old for purposes of determining fairness. An attempt to move an application to a queue will fail if the addition of the app's resources to that queue would violate the its maxRunningApps or maxResources constraints.