andygrove commented on issue #6165: URL: https://github.com/apache/datafusion-comet/issues/6165#issuecomment-5824273096
Two notes for when #6166 closes this. The perf table above undercounts the wrapper. The profile only resolved one of the two executors' copies of `libcomet`, so about half of the `__rust_alloc` and `__rust_dealloc` samples were missing. Summed over both executors, the wrapper was 7.6% of samples on the build this issue measured, not 4.9%. #6166's description has the corrected profile, and a re-run of all four builds in one session. #6166 doesn't add the `alloc_overhead` case from the Validation list. The benchmark links `comet` as an `rlib` into an executable, so on Linux it can't see the lookup. #6166 measured the change with the benchmark on macOS instead, where Mach-O goes through the TLV thunk even in an executable, and with the Q21 A/B and `perf` profile asked for above. The `dlopen` harness moved to #6213, which tracks removing the last lookup. That change trades the lookup for other per-call costs worth about a point of Q21 samples, so a per-call comparison through the real library matters more there. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
