[jira] [Commented] (ARROW-16014) [C++] Create more benchmarks for measuring expression evaluation overhead

["Aldrin M (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/ARROW-16014?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17512510#comment-17512510 ] 

Aldrin M commented on ARROW-16014:
----------------------------------

I wanted to consolidate a few references here as well from the user@ mailing list.

 

I originally thought that some of the compute API functions were leveraging parallelism when "enough data" was present. I posted on the user@ mailing list about this, with a [relevant repository][post-arrowuser-simplerepo] for reproducing the performance I was seeing.

 

In Weston's response, he noted [another post from Wes][post-arrowdev-execbatch] on the dev@ mailing list that seemed highly relevant.

 

[post-arrowuser-simplerepo]: [https://lists.apache.org/thread/blckonozxbtq0p4t9kvf087llsvvzsxq]

[post-arrowdev-execbatch] [https://lists.apache.org/thread/mp68ofm2hnvs2v2oz276rvw7y5kwqoyd]

 

> [C++] Create more benchmarks for measuring expression evaluation overhead
> -------------------------------------------------------------------------
>
>                 Key: ARROW-16014
>                 URL: https://issues.apache.org/jira/browse/ARROW-16014
>             Project: Apache Arrow
>          Issue Type: Improvement
>          Components: C++
>            Reporter: Weston Pace
>            Assignee: Weston Pace
>            Priority: Major
>
> We have a lot of great benchmarks on individual kernels but I don't know if we have enough benchmarks on the function & expression evaluation itself.
> There are some benchmarks (function_benchmark.cc) which measure this but I would like to get a better sense of a "bytes per second" number from the function system for a trivial function (e.g. an identity function that simply returns the values as is).
> In addition, we should measure overhead for common tasks like preallocation, etc.
> I would also like these benchmarks to be parameterized by batch size.  Being able to run on small batches enables (in theory) better cache utilization.  I suspect that the overhead here may start to become a bottleneck.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)