[infinispan-issues] [JBoss JIRA] (ISPN-7428) PutAll does not scale (triangle)

[Pedro Ruivo (JIRA)]

     [ https://issues.jboss.org/browse/ISPN-7428?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Pedro Ruivo updated ISPN-7428:
------------------------------
    Description: 
PutAll command for some reason now performs worse as the batch size grows.  There must be some exponential computation done now that needs to be cleaned up.

The new algorithm used for the triangle splits the command per segments. This slows down everything:

From
Before: 2150 iterations (430 per thread)
After: 104 iterations (20.8 per thread)

-Note:-
* -The per-segments are only need for primary>backup communication. It is possible to split by primary owner in the originator and the primary owner would order and split by segment-
(already implemented...)

  was:
PutAll command for some reason now performs worse as the batch size grows.  There must be some exponential computation done now that needs to be cleaned up.

The new algorithm used for the triangle splits the command per segments. This slows down everything:

From
Before: 2150 iterations (430 per thread)
After: 104 iterations (20.8 per thread)

-Note:
* The per-segments are only need for primary->backup communication. It is possible to split by primary owner in the originator and the primary owner would order and split by segment.- (already implemented...)



> PutAll does not scale (triangle)
> --------------------------------
>
>                 Key: ISPN-7428
>                 URL: https://issues.jboss.org/browse/ISPN-7428
>             Project: Infinispan
>          Issue Type: Bug
>          Components: Core
>            Reporter: William Burns
>            Assignee: Pedro Ruivo
>
> PutAll command for some reason now performs worse as the batch size grows.  There must be some exponential computation done now that needs to be cleaned up.
> The new algorithm used for the triangle splits the command per segments. This slows down everything:
> From
> Before: 2150 iterations (430 per thread)
> After: 104 iterations (20.8 per thread)
> -Note:-
> * -The per-segments are only need for primary>backup communication. It is possible to split by primary owner in the originator and the primary owner would order and split by segment-
> (already implemented...)



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