[CASSANDRA-13034] Move to FastThreadLocalThread and FastThreadLocal

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 3.11.0
• Severity: Normal
• Assignee: Robert Stupp
• Reporter: Robert Stupp

Authors

Robert Stupp

Reviewers

Jason Brown

• Created: 12/Dec/16 14:38
• Updated: 16/Apr/19 09:30
• Resolved: 13/Dec/16 16:53

Description

(Supersedes/includes CASSANDRA-13033 for 3.X & trunk)

We still use ThreadLocal in a couple of places, so I was curious how much faster FastThreadLocal is compared to ThreadLocal. A micro bench tells, that FastThreadLocal has a runtime of ~2.7ns and ThreadLocal of ~4.7ns - about 2ns slower (EDIT: subtracted baseline).

However, looking at the implementations it seems that ThreadLocal has more dependent pointer gets than FastThreadLocal. This (CPU cache misses) is not reflected in the artificial benchmark below.

The patch migrates all Thread instances (except a few in tests) and all ThreadLocal instances.

FastThreadLocalBench with 4 threads on 4 core CPU

     [java] FastThreadLocalBench.baseline                   2  avgt    5  3.023 ± 0.081  ns/op
     [java] FastThreadLocalBench.fastThreadLocal            2  avgt    5  5.610 ± 0.154  ns/op
     [java] FastThreadLocalBench.fastThreadLocal            4  avgt    5  5.653 ± 0.042  ns/op
     [java] FastThreadLocalBench.fastThreadLocal            8  avgt    5  5.763 ± 0.588  ns/op
     [java] FastThreadLocalBench.fastThreadLocal           12  avgt    5  5.673 ± 0.117  ns/op
     [java] FastThreadLocalBench.threadLocal                2  avgt    5  7.708 ± 0.723  ns/op
     [java] FastThreadLocalBench.threadLocal                4  avgt    5  7.604 ± 0.059  ns/op
     [java] FastThreadLocalBench.threadLocal                8  avgt    5  7.629 ± 0.080  ns/op
     [java] FastThreadLocalBench.threadLocal               12  avgt    5  7.858 ± 0.483  ns/op