[YARN-9509] Capped cpu usage with cgroup strict-resource-usage based on a mulitplier

• Type: New Feature
• Status: Open
• Priority: Minor
• Resolution: Unresolved
• Component/s: nodemanager
• Labels: pull-request-available
• Assignee: Unassigned
• Reporter: Nicolas Fraison
• Created: 24/Apr/19 12:17
• Updated: 03/Aug/21 14:01

Description

Add a multiplier configuration on strict resource usage to authorize container to use spare cpu up to a limit.
Currently with strict resource usage you can't get more than what you request which is sometime not good for jobs that doesn't have a constant usage of cpu (for ex. spark jobs with multiple stages).
But without strict resource usage we have seen some bad behaviour from our users that don't tune at all their needs and it leads to some containers requesting 2 vcore but constantly using 20.
The idea here is to still authorize containers to get more cpu than what they request if some are free but also to avoid too big differencies so SLA on jobs is not breached if the cluster is full (at least increase of runtime is contain)

Issue Links

• links to: GitHub Pull Request #766