Mailing-List: contact java-dev-help@lucene.apache.org; run by ezmlm
Precedence: bulk
Reply-To: java-dev@lucene.apache.org
Message-ID: <791040052.1213817265104.JavaMail.jira@brutus>
Date: Wed, 18 Jun 2008 12:27:45 -0700 (PDT)
From: "Rene Schwietzke (JIRA)" <jira@apache.org>
To: java-dev@lucene.apache.org
Subject: [jira] Commented: (LUCENE-1308) Remove String.intern() from
 Field.java to increase performance and lower contention
In-Reply-To: <1446841915.1213735964938.JavaMail.jira@brutus>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


    [ https://issues.apache.org/jira/browse/LUCENE-1308?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12606052#action_12606052 ] 

Rene Schwietzke commented on LUCENE-1308:
-----------------------------------------

A wrote a small test case that runs a single thread search, as well as a multithreaded search using the same indexsearcher. Especially when running in a threaded context, the replacement of String.intern() pays off. Even the single thread is faster.

I measured the following numbers:

String.Intern, Single Searcher
[main] Search  took: 3453ms
[Thread-2] Search  took: 17812ms
[Thread-3] Search  took: 18313ms
[Thread-1] Search  took: 18234ms
[Thread-0] Search  took: 18562ms

WeakHashMap, Single Searcher
[main] Search  took: 3156ms
[Thread-3] Search  took: 14953ms
[Thread-1] Search  took: 15593ms
[Thread-0] Search  took: 15656ms
[Thread-2] Search  took: 16188ms

ConcurrentHashMap, Single Searcher
[main] Search  took: 2844ms
[Thread-1] Search  took: 14812ms
[Thread-0] Search  took: 14890ms
[Thread-2] Search  took: 15172ms
[Thread-3] Search  took: 14656ms

> Remove String.intern() from Field.java to increase performance and lower contention
> -----------------------------------------------------------------------------------
>
>                 Key: LUCENE-1308
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1308
>             Project: Lucene - Java
>          Issue Type: Improvement
>    Affects Versions: 2.3.2
>            Reporter: Rene Schwietzke
>         Attachments: yad.zip
>
>
> Right now, *document.Field is interning all field names. While this makes sense because it lowers the overall memory consumption, the method intern() of String is know to be difficult to handle. 
> 1) it is a native call and therefore slower than anything on the Java level
> 2) the String pool is part of the perm space and not of the general heap, so it's size is more restricted and needs extra VM params to be managed
> 3) Some VMs show GC problems with strings in the string pool
> Suggested solution is a WeakHashMap instead, that takes care of unifying the String instances and at the same time keeping the pool in the heap space and releasing the String when it is not longer needed. For extra performance in a concurrent environment, a ConcurrentHashMap-like implementation of a weak hashmap is recommended, because we mostly read from the pool.
> We saw a 10% improvement in throughout and response time of our application and the application is not only doing searches (we read a lot of documents from the result). So a single measurement test case could show even more improvement in single and concurrent usage.
> The Cache:
> /** Cache to replace the expensive String.intern() call with the java version */
> private final static Map<String, WeakReference<String>> unifiedStringsCache =
>    Collections.synchronizedMap(new WeakHashMap<String, WeakReference<String>>(109));
> The access to it, instead of this.name = name.intern;
> // unify the strings, but do not use the expensive String.intern() version
> // which is not "weak enough", uses the perm space and is a native call
> String unifiedName = null;
> WeakReference<String> ref = unifiedStringsCache.get(name);
> if (ref != null)
> {
>     unifiedName = ref.get();
> }
> if (unifiedName == null)
> {
>     unifiedStringsCache.put(name, new WeakReference(name));
>     unifiedName = name;
> }
> this.name = unifiedName;
> I guess it is sufficient to have mostly all fields names interned, so I skipped the additional synchronization around the access and take the risk that only 99.99% :) of all field names are interned.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org