[ 
https://issues.apache.org/jira/browse/UIMA-6136?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16958009#comment-16958009
 ] 

Marshall Schor commented on UIMA-6136:
--------------------------------------

the traces show something strange.  From your description, I'm guessing that 
the index comparators are almost all equal.  Yet, it seems the get operation, 
which does first a hash, and then loops throught the bucket of equal hash hits, 
finds that bucket to be millions long.  

So, I think there's an actual bug somewhere in this implementation, which 
causes the cache.get to return null or a WeakReference which has been 
"cleared", causing it to over and over add the same index comparator to the 
cache.  

If you have any kind of a small test case :) ...

> FSIndexComparatorImpl.equalsWithoutType() gets slow for many CASes with the 
> same TS
> -----------------------------------------------------------------------------------
>
>                 Key: UIMA-6136
>                 URL: https://issues.apache.org/jira/browse/UIMA-6136
>             Project: UIMA
>          Issue Type: Bug
>          Components: UIMA
>    Affects Versions: 3.1.0SDK
>            Reporter: Richard Eckart de Castilho
>            Priority: Minor
>         Attachments: 2019-10-21_22-23-37.png, 2019-10-21_22-44-25.png
>
>
> When creating several hundred CASes with the same type system, the 
> `shareExisting` mechanism which is designed to save on memory starts eating 
> into CPU time quite a lot.
> This screenshot shows that in my particular case, the method is called ~11mio 
> times and takes the bulk of the processing time. The call hierarchy is a bit 
> messed up though - actually this happens when the CASes are initialized.
>  !2019-10-21_22-44-25.png|width=100%!
> The second screenshot shows the actual call hierarchy, but for some reason, 
> the profile doesn't properly dive into the `equals` method here and doesn't 
> count the time spent in `equalsWithoutType`.
>  !2019-10-21_22-23-37.png|width=100%!! 
> So either the method shouldn't be called that often - or - it should be way 
> faster.
> In the example, I have like 1800 CAS instances and their type system has 
> upwards of 200 types.



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

Reply via email to