zhengruifeng commented on pull request #29018: URL: https://github.com/apache/spark/pull/29018#issuecomment-657505193
I remove the specialization in methods and kryo registration, since they had ignorable impact on the performance. Then I retest on more config, and it seems that this regression comes from two parts: 1, a constant regression; 2, a bit regresion due to convert `Byte` to `Int` in training; Test Result: |Durations| GBT(depth=10;iter=10) | GBT(depth=10;iter=20) | RF(depth=10;iter=100) | RF(depth=20;iter=100) | |------|----------|------------|----------|------------| |This PR|559299|1028470|149084|956002| |Master|515832|997470|135395|942217| |Regression|8.4%|3.1%|10.1%|1.4%| When the computation of tree growth dominates the whole procedure, the regression （mainly due to part 2） can be tiny (<4%). ping @srowen ---------------------------------------------------------------- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
