JDK-8260940 : Regression ~3% in SPECjvm2008-Crypto.rsa-G1 in 17-b3
  • Type: Bug
  • Component: hotspot
  • Sub-Component: compiler
  • Affected Version: 17
  • Priority: P3
  • Status: Closed
  • Resolution: Won't Fix
  • Submitted: 2021-02-02
  • Updated: 2021-05-21
  • Resolved: 2021-05-21
Related Reports
Relates :  
Description
AFAICT this came in with CI build 65
Comments
We observed a huge increase in performance of Crypto.rsa with JDK-8262316 so we have decided this is not worth further work. If you agree you can close it as will not fix.
20-05-2021

System Configuration: Skylake Server ( Intel(R) Xeon(R) Gold 6139 CPU @ 2.30GHz) Command Line: numactl --cpunodebind=1 -l java -Xmx2g -Xmn2g -Xmn1g -XX:+UseParallelGC -jar SPECjvm2008.jar -ja "-XX:-UseFastStosb -XX:+UseXMMForObjInit -XX:-TieredCompilation -XX:UseAVX=2/3" -ikv -ict -coe -i 3 -it 15 -wt 30 -bt 1 crypto.rsa Scores: UseAVX=2 UseAVX=3 JDK-17-B3 : 545.98 ops/m 542.5 ops/m JDK-17-B2 : 544.34 ops/m 541.3 ops/m JDK-8257772 accelerated system memory initialization for constant sized allocations using AVX3 masked operations. It also changed the default initialization mechanism used by clear array to use vectorized initialization instead of fast stobs i.e. XMMForObjInit was made default true and UseFastStobs default false. Performance analysis does not show any major impact on SPECjvm2008 benchmark crypto.rsa, also there is usual 2-3% run-to-run variation in benchmark. Please provide further details or options to reproduce the issue.
19-05-2021

jdk-17+3-65 only contained JDK-8257772. Jatin, could you please have a look?
02-02-2021

ILW = Slight performance regression, with SPECjvm2008 benchmark, no known workaround = MMH = P3
02-02-2021