[
https://issues.apache.org/jira/browse/LUCENE-7882?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16599043#comment-16599043
]
Robert Muir commented on LUCENE-7882:
-------------------------------------
Well I think it needs to be tested on a modern JVM (e.g. latest release) to see
if that bug even still happens at all. And I don't agree that a cache certainly
makes it faster, we need to see real numbers to see if it matters. Historically
it has been far too difficult to remove such caches, e.g. scoring used to cache
sqrt() but then sqrt() got faster and the cache became unnecessary.
I don't want to see Accountable here as that won't help the issue. There is
more than just heap memory involved with compiled classes.
> Maybe expression compiler should cache recently compiled expressions?
> ---------------------------------------------------------------------
>
> Key: LUCENE-7882
> URL: https://issues.apache.org/jira/browse/LUCENE-7882
> Project: Lucene - Core
> Issue Type: Improvement
> Components: modules/expressions
> Reporter: Michael McCandless
> Priority: Major
>
> I've been running search performance tests using a simple expression
> ({{_score + ln(1000+unit_sales)}}) for sorting and hit this odd bottleneck:
> {noformat}
> "pool-1-thread-30" #70 prio=5 os_prio=0 tid=0x00007eea7000a000 nid=0x1ea8a
> waiting for monitor entry [0x00007eea867dd000]
> java.lang.Thread.State: BLOCKED (on object monitor)
> at
> org.apache.lucene.expressions.js.JavascriptCompiler$CompiledExpression.evaluate(_score
> + ln(1000+unit_sales))
> at
> org.apache.lucene.expressions.ExpressionFunctionValues.doubleValue(ExpressionFunctionValues.java:49)
> at
> com.amazon.lucene.OrderedVELeafCollector.collectInternal(OrderedVELeafCollector.java:123)
> at
> com.amazon.lucene.OrderedVELeafCollector.collect(OrderedVELeafCollector.java:108)
> at
> org.apache.lucene.search.MultiCollectorManager$Collectors$LeafCollectors.collect(MultiCollectorManager.java:102)
> at
> org.apache.lucene.search.Weight$DefaultBulkScorer.scoreAll(Weight.java:241)
> at
> org.apache.lucene.search.Weight$DefaultBulkScorer.score(Weight.java:184)
> at org.apache.lucene.search.BulkScorer.score(BulkScorer.java:39)
> at org.apache.lucene.search.IndexSearcher.search(IndexSearcher.java:658)
> at org.apache.lucene.search.IndexSearcher$5.call(IndexSearcher.java:600)
> at org.apache.lucene.search.IndexSearcher$5.call(IndexSearcher.java:597)
> at java.util.concurrent.FutureTask.run(FutureTask.java:266)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
> at java.lang.Thread.run(Thread.java:745)
> {noformat}
> I couldn't see any {{synchronized}} in the sources here, so I'm not sure
> which object monitor it's blocked on.
> I was accidentally compiling a new expression for every query, and that
> bottleneck would cause overall QPS to slow down drastically (~4X slower after
> ~1 hour of redline tests), as if the JVM is getting slower and slower to
> evaluate each expression the more expressions I had compiled.
> I tested JDK 9-ea and it also kept slowing down over time as the performance
> test ran.
> Maybe we should put a small cache in front of the expressions compiler to
> make it less trappy? Or maybe we can get to the root cause of why the JVM
> slows down more and more, the more expressions you compile?
> I won't have time to work on this in the near future so if anyone else feels
> the itch, please scratch it!
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
