[YARN-9116] Capacity Scheduler: implements queue level maximum-allocation inheritance

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.7.0
• Fix Version/s: 3.3.0
• Component/s: capacity scheduler
• Hadoop Flags: Reviewed

Release Note

After this change, capacity scheduler queue is able to inherit and override max-allocation from its parent queue. The subqueue's max-allocation can be larger or smaller than the parent queue's but can't exceed the global "yarn.scheduler.capacity.maximum-allocation". User can set queue-level max-allocation on any level of queues, using configuration property "yarn.scheduler.capacity.%QUEUE_PATH%.maximum-allocation". And this property allows user to set max-resource-allocation for customized resource types, e.g "memory=20G,vcores=20,gpu=3".

• Assignee: Aihua Xu
• Reporter: Aihua Xu
• Created: 12/Dec/18 21:58
• Updated: 24/Jan/19 17:57
• Resolved: 24/Jan/19 17:53

Description

YARN-1582 adds the support of maximum-allocation-mb configuration per queue which is targeting to support larger container features on dedicated queues (larger maximum-allocation-mb/maximum-allocation-vcores for such queue) . While to achieve larger container configuration, we need to increase the global maximum-allocation-mb/maximum-allocation-vcores (e.g. 120G/256) and then override those configurations with desired values on the queues since queue configuration can't be larger than cluster configuration. There are many queues in the system and if we forget to configure such values when adding a new queue, then such queue gets default 120G/256 which typically is not what we want.

We can come up with a queue-default configuration (set to normal queue configuration like 16G/8), so the leaf queues gets such values by default.