vertexclique commented on pull request #8598: URL: https://github.com/apache/arrow/pull/8598#issuecomment-725175419
So, I have fixed all unreproducible benchmarks that we are running in #8635. Also open-sourced a benchmark utility https://github.com/vertexclique/zor to create always the same baseline for all people who runs `cargo bench` somehow. Running master with reproducible benches and doing the same for this PR gives these results: https://gist.github.com/vertexclique/b5860ad836e78044d331cd3bb93fcf20 So far this PR is giving a boost to many operations. > I think we should address jhorstmann 's measurements of performance regressions before this pR is merged. I measured the performance. :upside_down_face: It is in the PR description. > I'm going to look at what the reason for the perf regressions is, during the week in my evenings. I haven't been able to profile criterion benchmarks with the tool that I use (https://superluminal.eu/), so I'll write some small application(s) so I can get the perf profiles for this. Would be nice! I have used the perf counters, and disasm so far and didn't see anything except bounds checks and proper alignment checks. Attached left is the master's code right is this pr. Mind that left doesn't jump to remainder checks because it is erased by const prop in remainder bits that I have applied with https://github.com/apache/arrow/commit/7696b89e68637ea1c54ef806a095a18ea7b2936f#diff-4eec3bf3d3993d5a6a7fa0b6a0b057dc5b517b7737eb47e00a887e7a3dcb1c37R65  > @vertexclique @jorgecarleitao there's been work on Java and C++ to support big-endian architectures; so maybe we can check in on previous mailing list discussions for guidance. I think CI might be the main concern (incl for arm-v7 support). Definitely, I will start a thread tomorrow. Also opened: https://github.com/apache/arrow/pull/8634 ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
