[BEAM-12127] Reduce counter overhead in PCollectionConsumerRegistry.accept

• Type: Bug
• Status: Resolved
• Priority: P2
• Resolution: Fixed
• Fix Version/s: Missing
• Component/s: java-fn-execution
• Assignee: Sam Whittle
• Reporter: Sam Whittle
• Created: 08/Apr/21 12:39
• Updated: 13/Jan/22 21:48
• Resolved: 13/Apr/21 06:47

Description

DelegatingCounter.inc shows up as 21% of cpu on nexmark query 2 benchmark under PCollectionConsumerRegistry.accept

2% is actual counter incrementing, but the majority is the delegation of DelegatingCounter which involves looking up thread-local state and then getting the counter for the name from the counter container. However in this case the counter container is known and can just be bound when constructing the counter instaed of using DelegatingCounter.

Issue Links

• links to: GitHub Pull Request #14473
• links to: GitHub Pull Request #14601