[
https://issues.apache.org/jira/browse/OPENNLP-421?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17798043#comment-17798043
]
ASF GitHub Bot commented on OPENNLP-421:
----------------------------------------
rzo1 commented on PR #568:
URL: https://github.com/apache/opennlp/pull/568#issuecomment-1859746321
Hi all,
I added related interner / dedup implementations inspired / based on the
code provided by @mawiesne from Aleksey.
I added the following interner impls:
- CHMStringInterner (as provided by Aleksey), thread-safe
- CHMStringDeduplicator (as provided by Aleksey in his talk), thread-safe ->
relaxes the canonical requirements on interning. It is more or less a
probabilistic deduplication
- HMStringInterner (as provided by Aleksey), not thread-safe
- JvmStringInterner -> relies on `String.intern()` - can be used to get the
previous OpenNLP behaviour ;)
- NoOpStringInterner -> doesn't actually intern/dedup Strings.
The implementation of the static `StringInterners` is based on how Hadoop is
doing it [1]. The default interner used in OpenNLP is now: `CHMStringInterner`
I added a system property `opennlp.interner.class`, which can be used to
specify the interner implementation which will be used at runtime:
- If people want the old behaviour back: they can.
- If people do not want interning at all: they can.
- If people want probabilistic dedup: they can.
Currently, an updated benchmark of the different impls is running as well as
a full eval build with the default.
Will update this PR with the JMH results in a few hours.
- [1]
https://github.com/c9n/hadoop/blob/master/hadoop-common-project%2Fhadoop-common%2Fsrc%2Fmain%2Fjava%2Forg%2Fapache%2Fhadoop%2Futil%2FStringInterner.java
> Large dictionaries cause JVM OutOfMemoryError: PermGen due to String interning
> ------------------------------------------------------------------------------
>
> Key: OPENNLP-421
> URL: https://issues.apache.org/jira/browse/OPENNLP-421
> Project: OpenNLP
> Issue Type: Bug
> Components: Name Finder
> Affects Versions: tools-1.5.2-incubating
> Environment: RedHat 5, JDK 1.6.0_29
> Reporter: Jay Hacker
> Assignee: Richard Zowalla
> Priority: Minor
> Labels: performance
> Original Estimate: 168h
> Remaining Estimate: 168h
>
> The current implementation of StringList:
> https://svn.apache.org/viewvc/incubator/opennlp/branches/opennlp-1.5.2-incubating/opennlp-tools/src/main/java/opennlp/tools/util/StringList.java?view=markup
>
> calls intern() on every String. Presumably this is an attempt to reduce
> memory usage for duplicate tokens. Interned Strings are stored in the JVM's
> permanent generation, which has a small fixed size (seems to be about 83 MB
> on modern 64-bit JVMs:
> [http://www.oracle.com/technetwork/java/javase/tech/vmoptions-jsp-140102.html]).
> Once this fills up, the JVM crashes with an OutOfMemoryError: PermGen
> space.
> The size of the PermGen can be increased with the -XX:MaxPermSize= option to
> the JVM. However, this option is non-standard and not well known, and it
> would be nice if OpenNLP worked out of the box without deep JVM tuning.
> This immediate problem could be fixed by simply not interning Strings.
> Looking at the Dictionary and DictionaryNameFinder code as a whole, however,
> there is a huge amount of room for performance improvement. Currently,
> DictionaryNameFinder.find works something like this:
> for every token in every tokenlist in the dictionary:
> copy it into a "meta dictionary" of single tokens
> for every possible subsequence of tokens in the sentence: // of which
> there are O(N^2)
> copy the sequence into a new array
> if the last token is in the "meta dictionary":
> make a StringList from the tokens
> look it up in the dictionary
> Dictionary itself is very heavyweight: it's a Set<StringListWrapper>, which
> wraps StringList, which wraps Array<String>. Every entry in the dictionary
> requires at least four allocated objects (in addition to the Strings): Array,
> StringList, StringListWrapper, and HashMap.Entry. Even contains and remove
> allocate new objects!
> From this comment in DictionaryNameFinder:
> // TODO: improve performance here
> It seems like improvements would be welcome. :) Removing some of the object
> overhead would more than make up for interning strings. Should I create a
> new Jira ticket to propose a more efficient design?
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
