[jira] [Comment Edited] (IGNITE-13086) Improve current page replacement mechanism.

["Stanilovsky Evgeny (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/IGNITE-13086?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17118406#comment-17118406 ] 

Stanilovsky Evgeny edited comment on IGNITE-13086 at 5/28/20, 7:34 AM:
-----------------------------------------------------------------------

jfr from new and baseline test execution attached, briefly - no allocation deviation found.
*replacement*.log is logs from IgnitePdsPageReplacementTestToYard.java from fix and baseline appropriate.
I also commited additional yardstick test [1] it shows performance boost, t11 - consist this fix, t12 - baseline.

 !screenshot-1.png! 

[1] https://issues.apache.org/jira/browse/IGNITE-13034 


was (Author: zstan):
jfr from new and baseline test execution attached, briefly - no фддщсфешщт deviation found.
*replacement*.log is logs from IgnitePdsPageReplacementTestToYard.java from fix and baseline appropriate.
I also commited additional yardstick test [1] it shows performance boost, t11 - consist this fix, t12 - baseline.

 !screenshot-1.png! 

[1] https://issues.apache.org/jira/browse/IGNITE-13034 

> Improve current page replacement mechanism.
> -------------------------------------------
>
>                 Key: IGNITE-13086
>                 URL: https://issues.apache.org/jira/browse/IGNITE-13086
>             Project: Ignite
>          Issue Type: Improvement
>          Components: persistence
>    Affects Versions: 2.8.1
>            Reporter: Stanilovsky Evgeny
>            Assignee: Stanilovsky Evgeny
>            Priority: Major
>         Attachments: 8.7-fix-replacement400_rand_512val_5touch_oldts.log, 8.7-replacement400_rand_512val_5touch_oldts.log, IgnitePdsPageReplacementTestToYard.java, replacement_64_new.jfr.zip, replacement_64_old.jfr.zip, screenshot-1.png
>
>          Time Spent: 10m
>  Remaining Estimate: 0h
>
> Experimentally proven that current page replacement functionality has problems with replace candidate computation. Current implementation obtain 5 random pages and make further decisions basing this pages last touch timestamp and some inner flags, however still possible cases when this pages set can be simply nullified due to inner logic. All improvements need to be proven, for example, by simple scenario: 
> 1. put some data until event EVT_PAGE_REPLACEMENT_STARTED is triggered
> 2. put 2 times more data than been loaded in p1.
> 3. execute fullscan (through ScanQuery) for old\cold data processing emulation.
> 4. start processing only pages which can fit into current mem region.
> 5. measure "replacedPages" metric.
> (i attach code mention above)



--
This message was sent by Atlassian Jira
(v8.3.4#803005)