Hi Pat, thanks a lot for this great proposal! I think it is very well structured and has the right level of detail. The improvements of your performance benchmarks look very promising and I think code-gen'd sorters would be a very nice improvement. I like that you plan to add a switch to activate this feature.
In order move on, we will need a committer who "champions" your FLIP, reviews the pull request, and eventually merges it. @Greg and @Stephan, what do you think about this proposal? Best, Fabian 2017-03-14 16:10 GMT+01:00 Pattarawat Chormai <pat.chor...@gmail.com>: > Hi all, > > I would like to initiate a discussion of applying code generation to > NormalizedKeySorter. The goal is to improve sorting performance > by generating suitable NormalizedKeySorter for underlying data. This > generated sorter will contains only necessary code in important methods, > such as swap and compare, hence improving sorting performance. > > Details of the implementation is illustrated at FLIP-18 : Code Generation > for improving sorting performance. > <https://cwiki.apache.org/confluence/display/FLINK/FLIP-18%3A+Code+Generation+for+improving+sorting+performance> > > > Also, because we’re doing it as a course project at TUB, we have > completed the implementation and made a pull-request > <https://github.com/apache/flink/pull/3511> to Flink repo already. > > From our evaluation, we have found that the pull-request reduces sorting > time around 7-10% and together with FLINK-3722 > <https://issues.apache.org/jira/browse/FLINK-3722> the sorting time is > decreased by 12-20%. > > > > Please take a look at the document and the pull-request and let me know if > you have any suggestion. > > Best, > Pat >
