felix-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pierre De Rop (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FELIX-5619) MultiProperyFilterIndex memory consumption
Date Mon, 29 May 2017 07:43:04 GMT

    [ https://issues.apache.org/jira/browse/FELIX-5619?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16028107#comment-16028107

Pierre De Rop commented on FELIX-5619:

Hi Jeroen,

I have committed a patch which allows you to supply your own FilterIndex implementation.
(notice that there was also an old jira issue (FELIX-4028) which was also asking for this).

A new org.apache.felix.dependencymanager.index.itest project has been created for this new
feature (and existing filter index tests that were previously provided in the org.apache.felix.dependencymanager.itest
bundle have been moved to org.apache.felix.dependencymanager.index.itest bundle).

So, you can supply a FilterIndex implementation in the two following ways:

1) statically: 

when you specify the "org.apache.felix.dependencymanager.filterindex" system property, you
can now optionally specify the classname which will be used for a given index.
The format is the following:

property-index ::= service-property | service-property ',' property-index
property-index-withclass ::=  classname ':' property-index
index ::= '*aspect*' | '*adapter*' | property-index | property-index-withclass
indices ::= index | indices ';' index

If you supply a classname, it must be part of a bundle which is a fragment of the org.apache.felix.dependencymanager



See the following tests for static custom filter indices:


2) dynamically:

You can also register in the OSGi service registry your own filter index (using or not dependency
manager API), and it will be registered in the ServiceRegistryCache.
In this case, you should start your bundle which register your FilterIndex (as well as the
org.apache.felix.dependencymanager bundle) first, before all other bundles. Else, previously
started dm component won't use your filter index.

See the following tests for dynamic custom filter indices:


> MultiProperyFilterIndex memory consumption
> ------------------------------------------
>                 Key: FELIX-5619
>                 URL: https://issues.apache.org/jira/browse/FELIX-5619
>             Project: Felix
>          Issue Type: Improvement
>          Components: Dependency Manager
>    Affects Versions: org.apache.felix.dependencymanager-r9
>            Reporter: Jeroen Daanen
>            Assignee: Pierre De Rop
>         Attachments: Filter.java, MultiPropertyFilterIndex.java, MultiPropertyFilterIndexPerformanceTest.java,
MultiPropertyFilterIndexReferencedTest.java, MultiPropertyKey.java, MultiPropertyKeyTest.java,
> In our product using the dependency manager and having configured several filter index
settings, the memory consumption of the MultiPropertyFilterIndex is relatively high because
it contains a map of string keys to service reference, where the keys are concatenated strings
based on the service properties. These strings are all newly created objects resulting in
inefficient memory usage.
> In the attachment you find a version of the index which uses less memory because it no
longer stores appended strings in new string objects, but stores the original strings in string
arrays. This way it can benefit for instance from the garbage collector deduplicating strings.

> Also attached are some unit tests to ensure the behavior did not change (note that we
could not find any existing unit tests on the MultiPropertyFilterIndex).
> If a large project is configured in our product we could have around 41500 services total.
The configured MultiproperyFilterIndexes then contain more then 500000 keys.
> When executing the attached unit test 'MultiPropertyFilterIndexPerformanceTest' for 500000
service references we found the memory consumption is reduced by 44%:
> Existing implemenation:
> fastest write time: 3.433 milliseconds
> fastest read time: 1.577 milliseconds  
> memory consumption: 1374 MB
> New implementation:
> fastest write time: 2.329 milliseconds
> fastest read time: 1.997 milliseconds  
> memory consumption: 771 MB
> Would it be either possible to replace the existing implementation or make it possible
to supply a custom implementation of a filter index to the dependency manager?

This message was sent by Atlassian JIRA

View raw message