On Thu, 17 Apr 2025 07:07:55 GMT, Radim Vansa <rva...@openjdk.org> wrote:
> This optimization is a followup to https://github.com/openjdk/jdk/pull/24290 > trying to reduce the performance regression in some scenarios introduced in > https://bugs.openjdk.org/browse/JDK-8292818 > > Iteration through the field stream of several variable-length-encoded items > is limited by inherent dependency of fully decoding the field before knowing > the position of next field. On the critical codepath addressed by this PR, > only the name and signature indices from are used though. > Inspired by the Varint-GB and Varint-G81U encodings, the idea of this change > is in adding a control stream to the stream of field infos; each field uses > one byte of this control stream, and currently the lower 6 bits encode the > length of the encoded field, allowing more efficient skipping through the > stream. The most significant bit duplicates the injected field flag > (name/signature lookup requires that), one bit is currently unused. > > My measurements on the attached reproducer > > hyperfine -w 50 -r 100 '/path/to/jdk-17/bin/java -cp /tmp CCC' > Benchmark 1: /path/to/jdk-17/bin/java -cp /tmp CCC > Time (mean ± σ): 51.3 ms ± 2.8 ms [User: 44.7 ms, System: 13.7 ms] > Range (min … max): 45.1 ms … 53.9 ms 100 runs > > > hyperfine -w 50 -r 100 '/path/to/jdk25-master/bin/java -cp /tmp CCC' > Benchmark 1: /path/to/jdk25-master/bin/java -cp /tmp CCC > Time (mean ± σ): 78.2 ms ± 1.0 ms [User: 74.6 ms, System: 17.3 ms] > Range (min … max): 73.8 ms … 79.7 ms 100 runs > > (the `jdk25-master` above already contains JDK-8353175) > > hyperfine -w 50 -r 100 '/path/to/jdk25-this-pr/bin/java -cp /tmp CCC' > Benchmark 1: /path/to/jdk25-this-pr/bin/java -cp /tmp CCC > Time (mean ± σ): 51.8 ms ± 2.1 ms [User: 48.9 ms, System: 16.8 ms] > Range (min … max): 47.4 ms … 55.1 ms 100 runs > > So in case of the synthetic reproducer we're already on the level of JDK 17. > However, the undisclosed production-grade reproducer still shows regression, > so there is still space for optimization. > > JDK 17: 1.6 s > JDK 21 (no patches): 22 s > JDK25-master: 12.3 s > JDK25-this-pr: 3.1 s > > > About the downsides: This PR increases the consumption by 1 byte for each > field in loaded classes. I've executed some tests on a simple Spring Boot > application with 6800 instance classes loaded -> about 16k fields in total, > with NMT on; the memory usage (Class.Metadata.used) seems to have grown by > 32kB; this discrepancy could be investigated later on. > > Architecturally, I had to remove `const`ness from some methods on ... This pull request has been closed without being integrated. ------------- PR: https://git.openjdk.org/jdk/pull/24713
