jmeter-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Felix Schumacher <felix.schumac...@internetallee.de>
Subject Re: Improve our percentiles computation
Date Wed, 15 May 2019 19:39:58 GMT

Am 15.05.19 um 17:28 schrieb Philippe Mouawad:
> Hello,
>
> I'll start with my team the development of a switch to better percentiles
> computation which will end as a PR contribution.
>
> To recap previous discussions:
>
>    - We have currently 2 implementations of Percentiles :
>       - A custom one used in old listeners
>       - A new one based on commons-math
>       - CONS:
>    - It impacts maintainability as a consequence
>       - The results are different which confuses users
>       - Percentiles are not explicitly documented regarding the algorithm
>       used which also creates confusion :
>          - https://github.com/apache/jmeter/pull/295
>          - https://github.com/apache/jmeter/pull/296
>       - Benefits:
>       - Have one algorithm in codebase
>       - Rely on best percentile algo for our use case
>       - Rely on performing and reliable library HdrHistogram/LatencyUtils
>       - Improve CO handling
>    - PREVIOUS ACTIONS:
>    - It was agreed that switching to HdrHistogram was a good option. This
>       library is a standard now:
>          - Used in MicroMeter (used by Spring)
>          - Used in future JMC
>          https://mail.openjdk.java.net/pipermail/jmc-dev/2019-May/000984.html
>       - Felix started a PR:
>          - https://github.com/apache/jmeter/pull/380/files

I updated the branch and thus the PR to be hopefully more correct than
the old one.

I still have a few doubts about the implementation. It seems to me, that
HdrHistogram is good for collecting the stats, but not optimized for
asking the stats a lot while the test is running.

That is the reason, that I tried to implement the calculation of some of
those stats myself. There is a good change, that I completely messed it up.

>       - Vladimir mentioned LatencyUtils in this PR:
>          - Used in MicroMeter
>          - Based on Gil Tene work who is master of CO related issues as
>          long as other titles
>
>
> *PROPOSED ACTION PLAN:*
>
>    - We'll start our work using LatencyUtils which seems interesting as it
>    provides a PauseDetector which will be useful in our case
>
>
> If you have any note on this, feel free to add your ideas in this thread.

Have a look at the performance of the reports while the stats are collected.

Felix

>
> Regards
> Philippe M.
> UbikLoadPack Team
>
>
>
>
>
>
> <https://www.openstreetmap.org/#map=18/50.69454/3.16455>
>

Mime
View raw message