[BEAM-11730] Reduce context switches for dataflow streaming appliance getdata reads

• Type: Bug
• Status: Resolved
• Priority: P2
• Resolution: Fixed
• Fix Version/s: Missing
• Component/s: runner-dataflow
• Assignee: Sam Whittle
• Reporter: Sam Whittle
• Created: 01/Feb/21 15:04
• Updated: 13/Jan/22 21:49
• Resolved: 12/Mar/21 08:25

Description

Currently reads are batched and issued from a separate thread pool of reading threads. However in the case that there is not read queuing due to the max # of parallel requests, switching threads is unnecessary and adds additional context swithing overhead.

In a benchmark I observed 5% cpu spent on LockSupport.park which some googling indicates is due to context switching.

Issue Links

• links to: GitHub Pull Request #13878