You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@flink.apache.org by Pattarawat Chormai <pa...@gmail.com> on 2017/04/04 00:46:39 UTC
Re: [DISCUSS] FLIP-18: Code Generation for improving sorting
performance
Hi guys,
I have made an additional optimization[1] related to megamorphic call issue
that Greg mentioned earlier. The optimization[2] improves execution time
around ~13%, while the original code from FLINK-5734 is ~11%.
IMHO, the improvement from metamorphic call optimization is very small
compared to the code we have to introduce. So, I think we can just go with
the PR that we currently have. What do you think?
[1]
https://github.com/heytitle/flink/commit/8e38b4d738b9953337361c62a8d77e909327d28f
[2]https://docs.google.com/spreadsheets/d/1PcdCdFX4bGecO6Lb5dLI2nww2NoeaA8c3MgbEdsVmk0/edit#gid=598217386
Best,
Pat
--
View this message in context: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-FLIP-18-Code-Generation-for-improving-sorting-performance-tp16486p16923.html
Sent from the Apache Flink Mailing List archive. mailing list archive at Nabble.com.
Re: [DISCUSS] FLIP-18: Code Generation for improving sorting performance
Posted by Pattarawat Chormai <pa...@gmail.com>.
Hi Greg,
Thanks for the input. Please let me know If you need some help. Meanwhile, I will look into code of the tasks that you mentioned.
Best,
Pat
> On Apr 5, 2017, at 5:42 PM, Greg Hogan <co...@greghogan.com> wrote:
>
> Pat,
>
> Thanks for running additional tests and continuing to work on this contribution.
>
> My testing is also showing that the performance gains remain even when multiple classes are used for sorting.
>
> I think we should proceed in the order of FLINK-3722, FLINK-4705, and FLINK-5734. Gabor has reviewed FLINK-3722 and I’ve done so multiple times. I’m looking into test coverage for FLINK-4705. Once these are reviewed and FLINK-5734 rebased we can benchmark Flink’s performance to validate the improvements.
>
> Greg
>
>
>> On Apr 3, 2017, at 8:46 PM, Pattarawat Chormai <pa...@gmail.com> wrote:
>>
>> Hi guys,
>>
>> I have made an additional optimization[1] related to megamorphic call issue
>> that Greg mentioned earlier. The optimization[2] improves execution time
>> around ~13%, while the original code from FLINK-5734 is ~11%.
>>
>> IMHO, the improvement from metamorphic call optimization is very small
>> compared to the code we have to introduce. So, I think we can just go with
>> the PR that we currently have. What do you think?
>>
>> [1]
>> https://github.com/heytitle/flink/commit/8e38b4d738b9953337361c62a8d77e909327d28f
>> [2]https://docs.google.com/spreadsheets/d/1PcdCdFX4bGecO6Lb5dLI2nww2NoeaA8c3MgbEdsVmk0/edit#gid=598217386
>>
>> Best,
>> Pat
>>
>>
>>
>> --
>> View this message in context: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-FLIP-18-Code-Generation-for-improving-sorting-performance-tp16486p16923.html
>> Sent from the Apache Flink Mailing List archive. mailing list archive at Nabble.com.
>
Re: [DISCUSS] FLIP-18: Code Generation for improving sorting
performance
Posted by Greg Hogan <co...@greghogan.com>.
Pat,
Thanks for running additional tests and continuing to work on this contribution.
My testing is also showing that the performance gains remain even when multiple classes are used for sorting.
I think we should proceed in the order of FLINK-3722, FLINK-4705, and FLINK-5734. Gabor has reviewed FLINK-3722 and I’ve done so multiple times. I’m looking into test coverage for FLINK-4705. Once these are reviewed and FLINK-5734 rebased we can benchmark Flink’s performance to validate the improvements.
Greg
> On Apr 3, 2017, at 8:46 PM, Pattarawat Chormai <pa...@gmail.com> wrote:
>
> Hi guys,
>
> I have made an additional optimization[1] related to megamorphic call issue
> that Greg mentioned earlier. The optimization[2] improves execution time
> around ~13%, while the original code from FLINK-5734 is ~11%.
>
> IMHO, the improvement from metamorphic call optimization is very small
> compared to the code we have to introduce. So, I think we can just go with
> the PR that we currently have. What do you think?
>
> [1]
> https://github.com/heytitle/flink/commit/8e38b4d738b9953337361c62a8d77e909327d28f
> [2]https://docs.google.com/spreadsheets/d/1PcdCdFX4bGecO6Lb5dLI2nww2NoeaA8c3MgbEdsVmk0/edit#gid=598217386
>
> Best,
> Pat
>
>
>
> --
> View this message in context: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-FLIP-18-Code-Generation-for-improving-sorting-performance-tp16486p16923.html
> Sent from the Apache Flink Mailing List archive. mailing list archive at Nabble.com.