Untangling Apache Hadoop YARN, Part 5: Using FairScheduler queue properties

Categories: Hadoop YARN

Previously in Part 4, we described the most commonly used FairScheduler properties in Apache Hadoop.  In Part 5, we’ll provide some examples to show how properties can be used, individually and in combination, to achieve commonly desired behavior such as application prioritization and organizing queues.

Example: Best Effort Queue

Summary: Create a “best effort” queue that runs applications when the cluster is underutilized.  

Implementation: In FairScheduler, a queue with weight 0.0 will only run applications if there is spare capacity in the cluster.  In other words, all the jobs in the priority_jobs queue will get allocated first, and then FairScheduler will allocate any such spare capacity to the best_effort_jobs queue.

Notes:

  • Jobs in cluster may take a long time to finish if the cluster is fully utilized with running jobs in other queues.

Caveat: There is a bug in early versions of CDH where the value 0.0 won’t quite work properly, but really tiny values (e.g. 0.01) work just fine.  YARN-5077 fixes the issue and is in releases CDH 5.7.3, CDH 5.8.2, CDH 5.9.0 and later.

Example: Using maxResources to Guarantee Resources for Low Latency Applications

Summary: Have a special queue which runs applications with special low latency requirements.

Implementation: Assume we have a cluster with the following resources: <memory: 20000 gb, vcores: 10000>.  By setting the maxResources property on the other_jobs queue, FairScheduler leaves <memory: 4000 gb, vcores: 2000> for the low_latency queue.

Notes:

  • All the applications in the other_jobs queue total utilization cannot exceed 80% of the cluster.
  • By leaving roughly 20% of the cluster for the low_latency queue, applications there can get started as quickly as possible.
  • This case is provided purely as an example.  In many cases, it will be preferable to use the “Queues for Low Latency Applications using Preemption” below.

Example: Using Preemption to Guarantee Resources to Low Latency Applications without Compromising on Utilization

Summary: Have a special queue which runs applications with special low latency requirements.

Implementation:  Assume we have a cluster with the following resources: <memory: 20000 gb, vcores: 10000>.  This configuration turns on preemption on the low_latency queue.

Notes:

  • Unlike the maxResources version (see previous example), the full cluster is available to the other_jobs queue, but applications in the other_jobs can be preempted in order for the low_latency queue to start applications running.
  • If you still wish to limit the total cluster usage of the low_latency queue, maxResources could be applied.
  • There are two more examples of priority queues later that show how preemption can be used in a more sophisticated fashion.
  • Reminder: To enable preemption in FairScheduler, this property must be set in yarn-site.xml:

Example: Limiting the Size of Ad-Hoc Queues

Summary: Allow child ad-hoc queues to have a maxResources setting

Implementation: Normally, it is impossible to set properties on ad-hoc queues, since they are not defined in the fair-scheduler.xml file.  By setting the maxChildResources property on the some_parent queue, any children of that queue (e.g. ad-hoc user queues or ad-hoc group queues) will have the equivalent of <maxResources>8192 mb,8 vcores</maxResources> set.

Notes:

  • This feature was introduced in and is new to CDH 5.9.0.

Example: Organizational Queues

Summary: Give each organization a queue for their applications.

Implementation: Provide a queue for each organization, in this case sales, marketing, finance, and data science.  Each group is given an equal share of Steady FairShare.

Hierarchical Implementation: The sales organization has northamerica and europe subqueues.  The marketing organization has reports and website queues.  The data_science organization has priority and best_effort queues.

Example: Strict Priority Queuing

Summary: This is an alternative approach to priority queues.

Implementation:   In the previous example, FairScheduler is using preemption to enforce a container allocation of 100/10/1.  In this version the root.other and root.other.other queues are given a weight of 0.  This has the following consequences:

  1. Any jobs in the priority1 queue will be fully allocated first, then any spare resources are given to jobs in the priority2 queue.  Finally, any spare resources after that will be given to the priority3 queue.
  2. If each of the priority queues have jobs beyond the capacity of the cluster, then jobs in the priority2 queue will only begin after the total resource requirements of all jobs in priority1 fall below the capacity of the cluster.  The same goes for jobs in the priority2 queue getting allocated ahead of jobs in the priority3 queue.
  3. If jobs are added to the priority1 queue, then containers will be allocated to those new jobs as tasks finish from the queues priority2 and priority3.  Similarly, if new jobs are added to the priority2 queue (and assuming priority1 jobs stay fully allocated), then those jobs will get containers as tasks finish in the priority3 queue.

Notes:

  • If preemption is enabled on all queues, then any jobs added to the priority1 queue will immediately preempt from jobs in priority2 and priority3.  Similarly, any added jobs in the priority2 queue will preempt from jobs in the priority3 queue.
  • As many levels of hierarchy can be added as needed in order to suit your needs.
  • As mentioned in the “Best Effort Queue” example earlier, there is a bug in early versions of CDH where the value 0.0 won’t quite work properly, but really tiny values (e.g. 0.01) work just fine.

Example: Implementing Priority Using Preemption

Summary: Given the following situation: (a) a cluster running at capacity, and (b) a “high priority” queue whose allocation is below its FairShare.  Turning on preemption guarantees that resources will be made available within the timeout value provided.

Implementation: Preemption must be enabled in FairScheduler by setting this property in yarn-site.xml:

Below is an example for a queue in fair-scheduler.xml.  If the queue does not receive 80% of its FairShare within 60 seconds, FairScheduler will begin preempting applications from some_other_queue and giving resources to priority_queue.

Example: Implementing More Levels of Priority Using Preemption

Summary: When two levels of priorities isn’t sufficient for separating job priority needs, you can have three levels of priorities with preemption.

Implementation: The high_priority and medium_priority queues have preemption enabled.  The low_priority queue will not have preemption enabled.  In order to prevent the medium_priority queue from preempting containers from the high_priority queue, we will also set the allowPreemptionFrom property on high_priority.

Notes: Notice this uses the new allowPreemptionFrom property, introduced in CDH 5.7.0.

Gotcha: Setting up a separate queue for Oozie launcher jobs

Summary: When Oozie starts an action, the launcher job requires using one MapReduce Application Master with one Task to actually launch the job.  In cases where Oozie launches many actions simultaneously, the queue and/or cluster could hit maxAMShare with Oozie Launcher MapReduce AMs, which will cause deadlock since any subsequent YARN applications cannot launch an AM.

The solution is to put the Oozie launcher jobs into a separate queue and restrict the queue as needed.

Example:

  • In your Oozie Workflows, set the property oozie.launcher.mapred.job.queue.name to point to the Oozie launcher queue:

  • And in FairScheduler, you should create the queue for the launcher jobs

Notes:

  • This is not a FairScheduler specific issue.  However, this interaction between Oozie and YARN is a sufficiently common issue that it is documented here.  This issue is especially common on small clusters.
  • Putting a restriction like maxResources or maxRunningApps on the launcher queue will help prevent the Oozie launcher jobs from deadlocking the cluster.

Gotcha: Setting hard limits on root or a parent queue

Summary: In FairScheduler, hard limits like maxRunningApps or maxResources propagate top-down.  Similarly, setting such a property on the root queue will affect all queues.

Example: root.parent1 has maxRunningApps set to 1.  As a result, despite the settings to maxRunningApps to a value greater than 1 in the childA and childB queues, only one application can be running in total.

Gotcha: Deleting Queues

Summary: There isn’t a direct way to delete queues by command line or UI.  However, if you update your fair-scheduler.xml configuration file and remove a queue that was in the file, then upon the next scheduler configuration refresh, the queue will be removed.

Solution: Make sure all applications in the queue have status FINISHED before actually changing the configuration.  In the case of MR applications, also make sure they are saved in the JobHistoryServer.

Gotcha: FairScheduler Preemption is Global

Summary: Preemption in FairScheduler is global.  The scheduler looks at all queues which are above their FairShare, takes those resources, and allocates them to queues which are below their FairShare.  There currently is no form of restricting preemption to a subset of queues or to only allow preemption from queueA to queueB.

Solution: The allowPreemptionFrom property allows some limited version of preemption control.  Future versions of FairScheduler may allow other forms of control or grouping.

Gotcha: Small Clusters and/or lots of Really Large Jobs

Summary: It is not uncommon for users to run small clusters, where number of active queues (with jobs) is comparable to the amount of resources (vcores or GBs) in the cluster. This can happen in testing/staging clusters. A similar situation ensues when too many large jobs are submitted to the same queue. In these cases, the cluster might get into a livelock or the jobs may be really slow.

Solution: Consider limiting the number of queues on a small cluster.  Also, consider limiting the number of running jobs in a queue by setting the maxRunningApps property.  For example:

Extra Reading

To get more information about Fair Scheduler, take a look at the online documentation (Apache Hadoop and Cloudera CDH 5.9.x versions are available).

Thanks

Thanks to Andre Araujo, Paul Battaglia, Stephanie Bodoff, Haibo Chen, Spandan Dutta, Yufei Gu, Jonathan Hsieh, Patrick Hunt, Adrian Kalaszi, Karthik Kambatla, Robert Kanter, Justin Kestelyn, Vijay Kalluru, Cris Morris, Sandy Ryza, Grant Sohn, Wilfred Spiegelenburg, Juan Yu, and Yongjun Zhang for their reviews and comments on this multi-part blog post.

Ray Chiang is a Software Engineer at Cloudera.
Dennis Dawson is a Senior Technical Writer at Cloudera.

 

facebooktwittergoogle_pluslinkedinmailfacebooktwittergoogle_pluslinkedinmail

Leave a Reply

Your email address will not be published. Required fields are marked *