[jboss-jira] [JBoss JIRA] (DROOLS-1494) Problems with parallel agenda

[Mario Fusco (JIRA)]

    [ https://issues.jboss.org/browse/DROOLS-1494?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13383681#comment-13383681 ] 

Mario Fusco commented on DROOLS-1494:
-------------------------------------

The reason why AdvOperators3ExpertBenchmark and RetractFactsFromWmExpertBenchmark are falling back to the single threaded agenda is because, while it's true that there are multiple partitions, each OTN has only one single partition below it. I corrected this behaviour and allowed parallel evaluation with this pull request: https://github.com/kiegroup/drools/pull/1163 However I ran some preliminary benchmark and found that in those case the parallel engine is slower than the sequential one, so I'd keep the fallback and drop that pull request.

[~tzimanyi] please also give a run at my pull request and let me know if you'll also find same results. In case you believe that the parallel engine is faster also in that case (I don't think so) feel free to merge my pull request.

Regarding the other 2 benchmarks that show no gain when running in parallel I believe we have a quite similar situation. There each OTN only has only 2 partitions below it (e.g. for constraints firstName == null and firstName != null) and then it is not able to fully leverage the 8-core cpu.

> Problems with parallel agenda
> -----------------------------
>
>                 Key: DROOLS-1494
>                 URL: https://issues.jboss.org/browse/DROOLS-1494
>             Project: Drools
>          Issue Type: Bug
>          Components: core engine
>    Affects Versions: 7.0.0.Beta7
>            Reporter: Tibor Zimányi
>            Assignee: Mario Fusco
>              Labels: reported-by-qe
>
> From doc [1] there can be seen that four benchmarks show problems when run with parallel agenda enabled. 
> Benchmarks _AdvOperators3ExpertBenchmark_ and _RetractFactsFromWmExpertBenchmark_ 
> falled back to single threaded agenda although they doesn't use features that aren't supported with parallel agenda (e.g. like salience). 
> Benchmarks StartsStartedbyFusionBenchmark and UpdateFactsInWmExpertBenchmark shows no gain when using parallel agenda although they contain more partitions. 
> I created runners for debugging here [2], so the benchmarks can be debugged outside of JMH. They are part of master branch in kie-benchmarks, so just pull of new changes is sufficient to get them. 
> [1] https://docs.google.com/a/redhat.com/spreadsheets/d/1hp5lLcmox_NFEUsqNSNZBW2XvkrJWgj0LtEiVvbsCYU/edit?usp=sharing
> [2] https://github.com/kiegroup/kie-benchmarks/commit/887f758931b02d096f1e870a3145119db315e1af



--
This message was sent by Atlassian JIRA
(v7.2.3#72005)