Performance Testing for JEP 450: Compact Object Headers

SPECjvm

This is SPECjvm2008 modified to run with JDK24. Scores are ops/min, more is better. All benchmarks have been run on a AWS c5.9xlarge instance.

JVM flags: -Xms8g -Xms8g -XX:+AlwaysPreTouch

Notes: The CryptoRsa benchmark shows a legit regression of ~2%. I confirmed this by running 50 times (most benchmarks are run 3 times). The regression disappears when improving oopDesc::klass().

Renaissance

This is renaissance-jmh-0.15.0, excluding Als, ChiSquare, DecTree, GaussMix, LogRegression, MovieLens, NaiveBayes, PageRank and DbShootout, which are not compatible with JDK24 due to SecurityManager restriction. Scores are ms/op, less is better. All benchmarks have been run on a AWS c5.9xlarge instance.

JVM flags: -Xms8g -Xms8g -XX:+AlwaysPreTouch

SPECjbb2015

SPECjbb2015 in composite mode, run on a AWS c5.9xlarge instance, 10 times in a row, scores averaged over all runs. More is better.

JVM flags: -Xms20g -Xms20g -XX:+AlwaysPreTouch

SPECjbb2015 (run at Red Hat)

The following tests were run at Red Hat on a bare metal RHEL9 x64 i7-4770 8core machine, with the benchmark process isolated on 6 cores, 10 consecutive runs, all scores averaged.

JVM flags: -Xshare:off -Xlog:gc* -XX:+UseG1GC -Xms4g -Xms4g

perf stat flags:  L1-dcache-load-misses,L1-dcache-loads,LLC-load-misses,LLC-loads,dTLB-load-misses,dTLB-loads,instructions,branches

To measure the reduction in liveset size, the same benchmark was repeated with a periodic Full GC being scheduled every 5 minutes, and the liveset size was measured after every full GC: