lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Paul Cowan (JIRA)" <>
Subject [jira] Updated: (LUCENE-806) Synchronization bottleneck in FieldSortedHitQueue with many concurrent readers
Date Sat, 17 Feb 2007 04:14:06 GMT


Paul Cowan updated LUCENE-806:

    Attachment: lucene-806.patch

Just to clarify, the specific issue here is that RuleBasedCollator (the only concrete implementation
of Collator in the JDK, and the one that is always returned by Collator.getInstance(Locale))
has a synchronized compare(), meaning that if you have many threads building FieldSortedHitQueues
with large results and locale-sensitive sorting, and they share a Collator, they end up waiting
for each other in that method (which can obviously be called tens of thousands of times during
a very large search). The way to get a Collator is by calling Collator.getInstance(Locale),
which makes it look like this problem is the JDK's fault; however, Collator.getInstance(Locale)
actually returns a clone() of the object from the cache. The caching mechanism seems to be
to prevent having to rebuld the rule tables, rather than the objects themselves. Therefore,
the JDK version balances performance and thread safety quite well. 

On the Lucene end, though, the FieldSortedHitQueue implements its own caching mechanism, meaning
that the generated ScoreDocComparators are cached (with no way to disable this behaviour,
even if you wanted to). Therefore, one you've got your comparator (the unique key being {reader,
fieldname, type, locale, factory}), every search using sorting on the same field in the same
way on the same reader will use the same Collator, possibly causing a synchronization bottleneck.
Even providong your own factory to the SortField doesn't REALLY help, as they're cached one
level 'above' that (you can work around this; see below)

Attached is a patch which provides a 'quick and dirty' way of dealing with this. NOTE: THIS
PATCH IS NOT PRODUCTION QUALITY, it's just a proof of concept. If people like the idea, I'll
tidy it up substantially.

This works by adding a flag, usePerThreadLocaleComparators, set by a static method, to FieldSortedHitQueue.
If this flag is NOT set, behaviour remains the same. If it's set to true, however, createValue
calls a per-thread version of comparatorStringLocale, which returns a simple wrapper ScoreDocComparator
which delegates to the one created by comparatorStringLocale, using a ThreadLocal to make
sure they're not shared between threads.

For demonstration purposes, I have added a quick demo main() method to FieldSortedHitQueue,
which does a simple timing test -- 20 threads each inserting 20000 dummy documents into a
200-element FieldSortedHitQueue. Note that it uses CountDownLatches to coordinate the threads,
so this dummy test will only run under Java 5. Sorry, but as a quick demo it will do for now.
By changing the values of
    final int threadCount = 20;
    final int docCount = 10000;
    final int queueSize = 200;
you can change the parameters I mentioned above. However, the figures seem to roughly the
same proportion regardless of how high or low those numbers are, within reason; the parameters
provided are  enough to spend a LOT of time waiting for locks; making them higher doesn't
really make that much difference. If anything, making the queuesize larger makes the new version
of the code (with the flag set) look better in comparison. On my dev machine (1.8 GHz Celeron
laptop) the test as-is gives the following figures:
this is a pretty substantial difference, and (I think) makes it worth pursuing this further.
Your mileage may vary, but between 2x-4x faster seems typical for anywhere above, say, 5 threads,
1000 docs, and queue size of 50.

So if people are happy for me to proceed down this path, I'm equally happy to tidy up and
produce a cleaner, documented etc. version of this patch. However, the more I look at this,
the more I'd like to refactor this code -- it's not the nicest code in Lucene, and I think
it could be tidied up (personally). My proposal would be something along the lines of changing
all those static methods in FieldSortedHitQueue (comparatorXXXX) to be implementations of
SortComparatorSource. There'd be a StringComparatorSource, AutoComparatorSource, etc. Everything
in FSHQ would be made to deal with the SortComparatorSources only, abstracting out all the
hard work. The logic of the create() method  could be replaced by a PerFieldComparatorSource,
which produces one or more of the others depending on the field type, much as it does now.
Everything else could be implemented (possibly) using the Decorator pattern to implement new
SortComparatorSources. Namely, a CachingComparatorSource would use some sort of caching mechanism
(possibly the FieldCacheImpl.Cache, as it does now, though that seems like an odd coupling)
to cache the SortComparatorSources produced by the PerFieldComparatorSource; then we're basically
back where we are now. Along comes the PerThreadComparatorSource, which uses ThreadLocals
to do basically what my patch above does. All these classes would be available externally,
for people to wrap around their own SortComparatorSources when setting up their SortFields;
if no factory is provided in the SortField, things work much as they do now.

What do people think? Is the quick and dirty way (a) worthwhile, and (b) good enough? Should
I look at implementing the bigger, fancier patch which will be more work and more complicated
but ultimately (I think) make FieldSortedHitQueues much cleaner? Or is there another alternative
(for example, another low-impact option would be doing something like what  I've done now,
but instead using a ThreadLocal in a Comparator implementation, which could mean that the
API for FieldSortedHitQueue doesn't need to change at all). Or is this not worth making part
of the source tree, given that there are ways around it (supplying your own SortComparatorSource
which manages its own ThreadLocals).  The performance gain IS substantial, though...

Apologies for length, this is quite a confusing area and I wanted to make sure I hadn't forgotten

> Synchronization bottleneck in FieldSortedHitQueue with many concurrent readers
> ------------------------------------------------------------------------------
>                 Key: LUCENE-806
>                 URL:
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Search
>    Affects Versions: 2.0.0
>            Reporter: Paul Cowan
>            Priority: Minor
>         Attachments: lucene-806.patch
> 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
>         - waiting to lock <0x6b1e8c68> (a java.text.RuleBasedCollator)
>         at$
>         at
>         at org.apache.lucene.util.PriorityQueue.upHeap(
>         at org.apache.lucene.util.PriorityQueue.put(
>         at org.apache.lucene.util.PriorityQueue.insert(
>         at
>         at
>         at
>         at
>         at
>         at
>         at
> .....
> In our case we had 12 threads waiting like this, while one thread had the lock on the
RuleBasedCollator.  Turns out RuleBasedCollator' 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:
For additional commands, e-mail:

View raw message