Repository: hadoop Updated Branches: refs/heads/branch-2.7 e160ddfb8 -> bade7f06e
YARN-4492. Add documentation for preemption supported in Capacity scheduler. Contributed by Naganarasimha G R Project: http://git-wip-us.apache.org/repos/asf/hadoop/repo Commit: http://git-wip-us.apache.org/repos/asf/hadoop/commit/bade7f06 Tree: http://git-wip-us.apache.org/repos/asf/hadoop/tree/bade7f06 Diff: http://git-wip-us.apache.org/repos/asf/hadoop/diff/bade7f06 Branch: refs/heads/branch-2.7 Commit: bade7f06e80ba7ca035c9b1b5d910acdc3cdbf19 Parents: e160ddf Author: Jason Lowe <jl...@apache.org> Authored: Thu Jan 21 15:43:23 2016 +0000 Committer: Jason Lowe <jl...@apache.org> Committed: Thu Jan 21 15:43:23 2016 +0000 ---------------------------------------------------------------------- hadoop-yarn-project/CHANGES.txt | 3 ++ .../src/site/markdown/CapacityScheduler.md | 31 +++++++++++++++++++- 2 files changed, 33 insertions(+), 1 deletion(-) ---------------------------------------------------------------------- http://git-wip-us.apache.org/repos/asf/hadoop/blob/bade7f06/hadoop-yarn-project/CHANGES.txt ---------------------------------------------------------------------- diff --git a/hadoop-yarn-project/CHANGES.txt b/hadoop-yarn-project/CHANGES.txt index b4fe340..e18b8e7 100644 --- a/hadoop-yarn-project/CHANGES.txt +++ b/hadoop-yarn-project/CHANGES.txt @@ -10,6 +10,9 @@ Release 2.7.3 - UNRELEASED YARN-4287. Capacity Scheduler: Rack Locality improvement. (Nathan Roberts via wangda) + YARN-4492. Add documentation for preemption supported in Capacity + scheduler (Naganarasimha G R via jlowe) + OPTIMIZATIONS BUG FIXES http://git-wip-us.apache.org/repos/asf/hadoop/blob/bade7f06/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-site/src/site/markdown/CapacityScheduler.md ---------------------------------------------------------------------- diff --git a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-site/src/site/markdown/CapacityScheduler.md b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-site/src/site/markdown/CapacityScheduler.md index 9227b1f..7b19acd 100644 --- a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-site/src/site/markdown/CapacityScheduler.md +++ b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-site/src/site/markdown/CapacityScheduler.md @@ -22,6 +22,7 @@ Hadoop: Capacity Scheduler * [Setting up `ResourceManager` to use `CapacityScheduler`](#Setting_up_ResourceManager_to_use_CapacityScheduler`) * [Setting up queues](#Setting_up_queues) * [Queue Properties](#Queue_Properties) + * [Capacity Scheduler container preemption](#Capacity_Scheduler_container_preemption) * [Other Properties](#Other_Properties) * [Reviewing the configuration of the CapacityScheduler](#Reviewing_the_configuration_of_the_CapacityScheduler) * [Changing Queue Configuration](#Changing_Queue_Configuration) @@ -57,7 +58,7 @@ The `CapacityScheduler` supports the following features: * **Security** - Each queue has strict ACLs which controls which users can submit applications to individual queues. Also, there are safe-guards to ensure that users cannot view and/or modify applications from other users. Also, per-queue and system administrator roles are supported. -* **Elasticity** - Free resources can be allocated to any queue beyond it's capacity. When there is demand for these resources from queues running below capacity at a future point in time, as tasks scheduled on these resources complete, they will be assigned to applications on queues running below the capacity (pre-emption is not supported). This ensures that resources are available in a predictable and elastic manner to queues, thus preventing artifical silos of resources in the cluster which helps utilization. +* **Elasticity** - Free resources can be allocated to any queue beyond its capacity. When there is demand for these resources from queues running below capacity at a future point in time, as tasks scheduled on these resources complete, they will be assigned to applications on queues running below the capacity (pre-emption is also supported). This ensures that resources are available in a predictable and elastic manner to queues, thus preventing artifical silos of resources in the cluster which helps utilization. * **Multi-tenancy** - Comprehensive set of limits are provided to prevent a single application, user and queue from monopolizing resources of the queue or the cluster as a whole to ensure that the cluster isn't overwhelmed. @@ -177,6 +178,34 @@ Example: </property> ``` + +### Capacity Scheduler container preemption + + The `CapacityScheduler` supports preemption of container from the queues whose resource usage is more than their guaranteed capacity. The following configuration parameters need to be enabled in yarn-site.xml for supporting preemption of application containers. + +| Property | Description | +|:---- |:---- | +| `yarn.resourcemanager.scheduler.monitor.enable` | Enable a set of periodic monitors (specified in yarn.resourcemanager.scheduler.monitor.policies) that affect the scheduler. Default value is false. | +| `yarn.resourcemanager.scheduler.monitor.policies` | The list of SchedulingEditPolicy classes that interact with the scheduler. Configured policies need to be compatible with the scheduler. Default value is `org.apache.hadoop.yarn.server.resourcemanager.monitor.capacity.ProportionalCapacityPreemptionPolicy` which is compatible with `CapacityScheduler` | + +The following configuration parameters can be configured in yarn-site.xml to control the preemption of containers when `ProportionalCapacityPreemptionPolicy` class is configured for `yarn.resourcemanager.scheduler.monitor.policies` + +| Property | Description | +|:---- |:---- | +| `yarn.resourcemanager.monitor.capacity.preemption.observe_only` | If true, run the policy but do not affect the cluster with preemption and kill events. Default value is false | +| `yarn.resourcemanager.monitor.capacity.preemption.monitoring_interval` | Time in milliseconds between invocations of this ProportionalCapacityPreemptionPolicy policy. Default value is 3000 | +| `yarn.resourcemanager.monitor.capacity.preemption.max_wait_before_kill` | Time in milliseconds between requesting a preemption from an application and killing the container. Default value is 15000 | +| `yarn.resourcemanager.monitor.capacity.preemption.total_preemption_per_round` | Maximum percentage of resources preempted in a single round. By controlling this value one can throttle the pace at which containers are reclaimed from the cluster. After computing the total desired preemption, the policy scales it back within this limit. Default value is `0.1` | +| `yarn.resourcemanager.monitor.capacity.preemption.max_ignored_over_capacity` | Maximum amount of resources above the target capacity ignored for preemption. This defines a deadzone around the target capacity that helps prevent thrashing and oscillations around the computed target balance. High values would slow the time to capacity and (absent natural.completions) it might prevent convergence to guaranteed capacity. Default value is `0.1` | +| `yarn.resourcemanager.monitor.capacity.preemption.natural_termination_factor` | Given a computed preemption target, account for containers naturally expiring and preempt only this percentage of the delta. This determines the rate of geometric convergence into the deadzone (`MAX_IGNORED_OVER_CAPACITY`). For example, a termination factor of 0.5 will reclaim almost 95% of resources within 5 * #`WAIT_TIME_BEFORE_KILL`, even absent natural termination. Default value is `0.2` | + + The `CapacityScheduler` supports the following configurations in capacity-scheduler.xml to control the preemption of application containers submitted to a queue. + +| Property | Description | +|:---- |:---- | +| `yarn.scheduler.capacity.<queue-path>.disable_preemption` | This configuration can be set to `true` to selectively disable preemption of application containers submitted to a given queue. This property applies only when system wide preemption is enabled by configuring `yarn.resourcemanager.scheduler.monitor.enable` to *true* and `yarn.resourcemanager.scheduler.monitor.policies` to *ProportionalCapacityPreemptionPolicy*. If this property is not set for a queue, then the property value is inherited from the queue's parent. Default value is false. + + ###Other Properties * Resource Calculator