lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From TCK <moonwatcher32...@gmail.com>
Subject Re: heap memory issues when sorting by a string field
Date Tue, 08 Dec 2009 06:45:34 GMT
Thanks for the feedback guys. The evidence I have collected does point to an
issue either in the java WeakHashMap implementation or in Lucene's use of
it. In particular, I used reflection to replace the WeakHashMap instances
with my own dummy Map that does a no-op for the put operation, and although
this caused tons more garbage to be created for each search query the CMS
executions always collected that garbage and brought down the tenured space
usage to a small value.

Before that, I tried calling clear() on the WeakHashMap instances and then
calling size() which calls expungeStaleEntries() but that didn't help. There
may be indeed be a case of double memory usage for a while (until enough CMS
executions have run) as Tom pointed out, but I don't fully understand what's
happenning yet.

Also, bizzarely during my introspection of the readerCache WeakHashMap
instances I found an instance of
org.apache.lucene.index.CompoundFileReader$CSIndexInput as a key. Looking at
the code I don't see how anything other than an IndexReader could possibly
get there, and this class certainly doesn't extend IndexReader. Any ideas?
:-) Maybe I need to get more sleep.

Btw, is searching with a sort by a text field not a common use-case of
lucene? I've been testing with only 1Gb indexes and I'm pretty sure there
are much larger indexes out there.

Cheers,
TCK





On Mon, Dec 7, 2009 at 7:57 PM, Tom Hill <solr-list@worldware.com> wrote:

> Hey, that's a nice little Class! I hadn't see it before. But it sounds like
> the asynchronous cleanup might deal with the problem I mentioned above (but
> I haven't looked at the code yet).
>
> It's an apache license - but you mentioned something about no third party
> libraries. Is that a policy for Lucene?
>
> Thanks,
>
> Tom
>
>
> On Mon, Dec 7, 2009 at 4:44 PM, Jason Rutherglen <
> jason.rutherglen@gmail.com
> > wrote:
>
> > I wonder if Google Collections (even though we don't use third party
> > libraries) concurrent map, which supports weak keys, handles the
> > removal of weakly referenced keys in a more elegant way than Java's
> > WeakHashMap?
> >
> > On Mon, Dec 7, 2009 at 4:38 PM, Tom Hill <solr-list@worldware.com>
> wrote:
> > > Hi -
> > >
> > > If I understand correctly, WeakHashMap does not free the memory for the
> > > value (cached data) when the key is nulled, or even when the key is
> > garbage
> > > collected.
> > >
> > > It requires one more step: a method on WeakHashMap must be called to
> > allow
> > > it to release its hard reference to the cached data. It appears that
> most
> > > methods in WeakHashMap end up calling expungeStaleEntries, which will
> > clear
> > > the hard reference. But you have to call some method on the map, before
> > the
> > > memory is eligible for garbage collection.
> > >
> > > So it requires four stages to free the cached data. Null the key; A GC
> to
> > > release the weak reference to the key; A call to some method on the
> map;
> > > Then the next GC cycle should free the value.
> > >
> > > So it seems possible that you could end up with double memory usage for
> a
> > > time. If you don't have a GC between the time that you close the old
> > reader,
> > > and you start to load the field cache entry for the next reader, then
> the
> > > key may still be hanging around uncollected.
> > >
> > > At that point, it may run a GC when you allocate the new cache, but
> > that's
> > > only the first GC. It can't free the cached data until after the next
> > call
> > > to expungeStaleEntries, so for a while you have both caches around.
> > >
> > > This extra usage could cause things to move into tenured space. Could
> > this
> > > be causing your problem?
> > >
> > > Workaround would be to cause some method to be called on the
> WeakHashMap.
> > > You don't want to call get(), since that will try to populate the
> cache.
> > > Maybe if you tried putting a small value to the cache, and doing a GC,
> > and
> > > see if your memory drops then.
> > >
> > >
> > > Tom
> > >
> > >
> > >
> > > On Mon, Dec 7, 2009 at 1:48 PM, TCK <moonwatcher32329@gmail.com>
> wrote:
> > >
> > >> Thanks for the response. But I'm definitely calling close() on the old
> > >> reader and opening a new one (not using reopen). Also, to simplify the
> > >> analysis, I did my test with a single-threaded requester to eliminate
> > any
> > >> concurrency issues.
> > >>
> > >> I'm doing:
> > >> sSearcher.getIndexReader().close();
> > >> sSearcher.close(); // this actually seems to be a no-op
> > >> IndexReader newIndexReader = IndexReader.open(newDirectory);
> > >> sSearcher = new IndexSearcher(newIndexReader);
> > >>
> > >> Btw, isn't it bad practice anyway to have an unbounded cache? Are
> there
> > any
> > >> plans to replace the HashMaps used for the innerCaches with an actual
> > >> size-bounded cache with some eviction policy (perhaps EhCache or
> > something)
> > >> ?
> > >>
> > >> Thanks again,
> > >> TCK
> > >>
> > >>
> > >>
> > >>
> > >> On Mon, Dec 7, 2009 at 4:37 PM, Erick Erickson <
> erickerickson@gmail.com
> > >> >wrote:
> > >>
> > >> > What this sounds like is that you're not really closing your
> > >> > readers even though you think you are. Sorting indeed uses up
> > >> > significant memory when it populates internal caches and keeps
> > >> > it around for later use (which is one of the reasons that warming
> > >> > queries matter). But if you really do close the reader, I'm pretty
> > >> > sure the memory should be GC-able.
> > >> >
> > >> > One thing that trips people up is IndexReader.reopen(). If it
> > >> > returns a reader different than the original, you *must* close the
> > >> > old one. If you don't, the old reader is still hanging around and
> > >> > memory won't be returne.... An example from the Javadocs...
> > >> >
> > >> >  IndexReader reader = ...
> > >> >  ...
> > >> >  IndexReader new = r.reopen();
> > >> >  if (new != reader) {
> > >> >   ...     // reader was reopened
> > >> >   reader.close();
> > >> >  }
> > >> >  reader = new;
> > >> >  ...
> > >> >
> > >> >
> > >> > If this is irrelevant, could you post your close/open
> > >> >
> > >> > code?
> > >> >
> > >> > HTH
> > >> >
> > >> > Erick
> > >> >
> > >> >
> > >> > On Mon, Dec 7, 2009 at 4:27 PM, TCK <moonwatcher32329@gmail.com>
> > wrote:
> > >> >
> > >> > > Hi,
> > >> > > I'm having heap memory issues when I do lucene queries involving
> > >> sorting
> > >> > by
> > >> > > a string field. Such queries seem to load a lot of data in to
the
> > heap.
> > >> > > Moreover lucene seems to hold on to references to this data even
> > after
> > >> > the
> > >> > > index reader has been closed and a full GC has been run. Some
of
> the
> > >> > > consequences of this are that in my generational heap
> configuration
> > a
> > >> lot
> > >> > > of
> > >> > > memory gets promoted to tenured space each time I close the old
> > index
> > >> > > reader
> > >> > > and after opening and querying using a new one, and the tenured
> > space
> > >> > > eventually gets fragmented causing a lot of promotion failures
> > >> resulting
> > >> > in
> > >> > > jvm hangs while the jvm does stop-the-world GCs.
> > >> > >
> > >> > > Does anyone know any workarounds to avoid these memory issues
when
> > >> doing
> > >> > > such lucene queries?
> > >> > >
> > >> > > My profiling showed that even after a full GC lucene is holding
on
> > to a
> > >> > lot
> > >> > > of references to field value data notably via the
> > >> > > FieldCacheImpl/ExtendedFieldCacheImpl. I noticed that the
> > WeakHashMap
> > >> > > readerCaches are using unbounded HashMaps as the innerCaches
and I
> > used
> > >> > > reflection to replace these innerCaches with dummy empty HashMaps,
> > but
> > >> > > still
> > >> > > I'm seeing the same behavior. I wondered if anyone has gone
> through
> > >> these
> > >> > > same issues before and would offer any advice.
> > >> > >
> > >> > > Thanks a lot,
> > >> > > TCK
> > >> > >
> > >> >
> > >>
> > >
> >
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
> > For additional commands, e-mail: java-user-help@lucene.apache.org
> >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message