[
https://issues.apache.org/jira/browse/SLING-12217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18073118#comment-18073118
]
Joerg Hoh commented on SLING-12217:
-----------------------------------
Yes, just looking at the code, {{isModelClass(Class<?> type)}} has a boolean
return value, and in that case one can assume that both {{true}} and {{false}}
are somehow distributed equally. I don't have data for it and rely on gut
feeling here, but I doubt that {{loadObject}} is frequently invoked for a
class, for which no model exists.
And that means, that {{isModelClass(Class<?> type)}} returns much more
frequently "true" than "false", and for that case an optimization for the miss
is hardly helpful.
Feel free to instrument that code and validate locally on a test system how
often that case miss actually happens.
But in any case: These 2 optimizations do not conflict at all, so we could
incorporate them both. I am just unwilling to add much code complexity for an
performance optimization in a code path, which is barely executed.
> Reduce resource lookups in Sling Model resolution
> -------------------------------------------------
>
> Key: SLING-12217
> URL: https://issues.apache.org/jira/browse/SLING-12217
> Project: Sling
> Issue Type: Task
> Components: Sling Models
> Affects Versions: Models Implementation 1.6.4
> Reporter: Joerg Hoh
> Priority: Major
>
> While analyzing repository access, I found a pattern in the Scripting HTL
> implementation, which leads to multiple invocations of the same check on the
> same resources in the Sling Model ModelAdapterFactory.
> Invocation at
> [JavaUseProvider.loadObject()|https://github.com/apache/sling-org-apache-sling-scripting-sightly/blob/690a818692bd08c3f6a49e842ea530cc3e60e3ad/src/main/java/org/apache/sling/scripting/sightly/impl/engine/extension/use/JavaUseProvider.java#L172]:
> * modelAdapterFactory.canCreateModelFromAdaptable()
> * modelAdapterFactory.createModel
> those methods directly or indirectly call
> [org.apache.sling.models.impl.AdapterImplementations.lookup()|https://github.com/apache/sling-org-apache-sling-models-impl/blob/cf088713c402177b9d96a5229567804510ef9918/src/main/java/org/apache/sling/models/impl/AdapterImplementations.java#L192],
> which calls
> [getModelClassForResource|https://github.com/apache/sling-org-apache-sling-models-impl/blob/cf088713c402177b9d96a5229567804510ef9918/src/main/java/org/apache/sling/models/impl/AdapterImplementations.java#L302];
> and there multiple resources are resolved, based on the provided
> resourceType.
> I benchmarked this code using a simple AEM sample page (WKND) using HTL and
> Sling Models. I found this code is responsible to the creation of 712 JCR
> Node resource objects (the total page rendering triggered the creation of
> more than 10k, so these are around 7%). When deduplicating the requested
> paths, I found that only 36 distinct paths were resolved with these 712 JCR
> Node resources.
> That means that with proper caching (probably using the
> ResourceResolver.getPropertyMap) around 6.5% of repository access could be
> avoided, leading to a page rendering improvement in about the same range.
>
>
>
>
>
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
