Synchronization bottleneck in FieldSortedHitQueue with many concurrent readers
------------------------------------------------------------------------------
Key: LUCENE-806
URL: https://issues.apache.org/jira/browse/LUCENE-806
Project: Lucene - Java
Issue Type: Improvement
Components: Search
Affects Versions: 2.0.0
Reporter: Paul Cowan
Priority: Minor
The below is from a post by (my colleague) Paul Smith to the java-users list:
---
Hi ho peoples.
We have an application that is internationalized, and stores data from many languages (each
project has it's own index, mostly aligned with a single language, maybe 2).
Anyway, I've noticed during some thread dumps diagnosing some performance issues, that there
appears to be a _potential_ synchronization bottleneck using Locale-based sorting of Strings.
I don't think this problem is the root cause of our performance problem, but I thought I'd
mention it here. Here's the stack dump of a thread waiting:
"http-1001-Processor245" daemon prio=1 tid=0x31434da0 nid=0x3744 waiting for monitor entry
[0x2cd44000..0x2cd45f30]
at java.text.RuleBasedCollator.compare(RuleBasedCollator.java)
- waiting to lock <0x6b1e8c68> (a java.text.RuleBasedCollator)
at org.apache.lucene.search.FieldSortedHitQueue$4.compare(FieldSortedHitQueue.java:320)
at org.apache.lucene.search.FieldSortedHitQueue.lessThan(FieldSortedHitQueue.java:114)
at org.apache.lucene.util.PriorityQueue.upHeap(PriorityQueue.java:120)
at org.apache.lucene.util.PriorityQueue.put(PriorityQueue.java:47)
at org.apache.lucene.util.PriorityQueue.insert(PriorityQueue.java:58)
at org.apache.lucene.search.FieldSortedHitQueue.insert(FieldSortedHitQueue.java:90)
at org.apache.lucene.search.FieldSortedHitQueue.insert(FieldSortedHitQueue.java:97)
at org.apache.lucene.search.TopFieldDocCollector.collect(TopFieldDocCollector.java:47)
at org.apache.lucene.search.BooleanScorer2.score(BooleanScorer2.java:291)
at org.apache.lucene.search.IndexSearcher.search(IndexSearcher.java:132)
at org.apache.lucene.search.IndexSearcher.search(IndexSearcher.java:110)
at com.aconex.index.search.FastLocaleSortIndexSearcher.search(FastLocaleSortIndexSearcher.java:90)
.....
In our case we had 12 threads waiting like this, while one thread had the lock on the RuleBasedCollator.
Turns out RuleBasedCollator's.compare(...) method is synchronized. I wonder if a ThreadLocal
based collator would be better here... ? There doesn't appear to be a reason for other threads
searching the same index to wait on this sort. Be just as easy to use their own. (Is RuleBasedCollator
a "heavy" object memory wise? Wouldn't have thought so, per thread)
Thoughts?
---
I've investigated this somewhat, and agree that this is a potential problem with a series
of possible workarounds. Further discussion (including proof-of-concept patch) to follow.
--
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
|