[BEAM-12402] Optimize PCollectionConsumerRegistry$MultiplexingMetricTrackingFnDataReceiver

• Type: Bug
• Status: Resolved
• Priority: P2
• Resolution: Fixed
• Fix Version/s: Missing
• Component/s: sdk-java-harness
• Assignee: Sam Whittle
• Reporter: Sam Whittle
• Created: 25/May/21 13:33
• Updated: 13/Jan/22 21:49
• Resolved: 25/May/21 21:16

Description

In Nexmark benchmark profile this was using 2% of cpu on
AbstractMapBasedMultimap$WrappedCollection$WrappedIterator
methods.

I believe this is due to the list returned here being a wrapper around the multiset.
https://github.com/apache/beam/blob/8463a054c1d7e2b7ee8d11e9569e065cb5e02196/sdks/java/harness/src/main/java/org/apache/beam/fn/harness/data/PCollectionConsumerRegistry.java#L172

We iterate over this list many times for counters. It appears that we don't need a wrapped list as it is documented that all consumers should be registered first. So it seems we can just create a copy there to an immutable list and trivially save that cpu.

Issue Links

• links to: GitHub Pull Request #14879