JDK-8321302 : GenShen: adapative tenuring: xeno down
  • Type: Sub-task
  • Component: hotspot
  • Sub-Component: gc
  • Priority: P4
  • Status: Open
  • Resolution: Unresolved
  • Submitted: 2023-12-04
  • Updated: 2024-01-17
Description
This is the first tweak listed in the parent ticket and includes the following changes:

1. (flip default to false) don't ignore mortality rates of cohorts older than previous tenuring threshold plus 1.

2. When the newly calculated tenuring threshold based solely on mortality rate as computed by the current algorithm is less than the tenuring threshold used in the previous epoch, then "xeno down" to the target new threshold by going halfway down (rounded down to an integer value). This makes the lowering of the tenuring threshold proceed more slowly than was previously the case.

Performance numbers will be included below.
Comments
Suspending for now. The work in this task has been completed but showed no appreciable performance improvement. The work is being suspended for now and will be picked up after GenShen separation work progresses further.
17-01-2024

Diff/Delta on github: https://github.com/openjdk/shenandoah/compare/master...ysramakrishna:shenandoah:better_tenuring
07-12-2023

Perf numbers in progress. Some early results below: 1. the initial impression that there was an improvement with this change on Kelvin's extremem workload turned out to be false. 2. More extensive testing revealed that while the mean was not affected, the run-to-run variability in the number of young collections went up with these changes. This was also the case with the benchmark's customer preparation time distribution where the average of the distributions across runs stayed about the same but there were run-to-run variations in these. I am running some numbers with specjbb as well as planning to look at the numbers from codepipeline testing, but although this seemed conceptually to be an approach that held promise, the performance figures do not bear this out, at least not on its own at this point.
07-12-2023