lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Robert Engels" <reng...@ix.netcom.com>
Subject RE: [jira] Updated: (LUCENE-529) TermInfosReader and other + instance ThreadLocal => transient/odd memory leaks => OutOfMemoryException
Date Thu, 23 Mar 2006 18:49:45 GMT
The testcase is still not correct - at least with regards to Lucene.

Review the ThreadLocal and ThreadLocalMap code again, you will see that
references to the ThreadLocal are kept using weak references, in a slot in
an array, and entries are reclaimed() (i.e. the slot cleared) PERIODICALLY
as new entries are ADDED or RETRIEVED. The array will never decrease in
size, but it is unlikely to grow to be very large.

See TheadLocalMap.cleanSomeSlots() and where/when it is called.

**** The reason that your test fails is that each entry in the table
maintains a HARD reference to the value, and since the entries are only
reclaimed PERIODICALLY (for performance reasons - see cleanSomeSlots()), you
do have the possibility to run out of memory (since the ThreadLocal values
you are storing are EXTREMELY LARGE).

In Lucene, the values used in the ThreadLocal are 1/100000 the size, and
thus should cause no problems.

-----Original Message-----
From: Andy Hind [mailto:andy.hind@alfresco.org]
Sent: Thursday, March 23, 2006 12:17 PM
To: java-dev@lucene.apache.org; rengels@ix.netcom.com
Subject: RE: [jira] Updated: (LUCENE-529) TermInfosReader and other +
instance ThreadLocal => transient/odd memory leaks =>
OutOfMemoryException



Well, unfortunately, it is your test case that is not equivalent.

OK you make 10 times as many objects that are 1/10 the size.
But the thread local map is the same size and likely to end up holding
the same number of stale entries, so yes the memory footprint is 10
times smaller and it works.

If you increase the size of the thread local table by x10 it will fail
again. This will not be an exact memory size match due to the underlying
algorithm for cleaning stuff out looking at log2(current.no.of thread
locals) + ?*log2(table.size) and being dependent on GC, so I have no
idea what fraction of stale entries will remain for any size.

I suggest you look at the source code for ThreadLocal and how stale
entries are removed.... look for "GC" and you will see where the "value"
is set to null to help the GC.

I should have proposed using SoftReferences not WeakReferences.
Apologies. Soft References also work fine.

The finalize method is indeed pointless but makes no difference to the
result. The point I was making is that it is pointless in lucene.

Yes, creating threads is expensive - my point - somewhat tongue-in-cheek
is at least they would be clean :-)

You can add an assert to make sure the weak/soft reference exists after
the Object is created. It does. You *may* be unlucky and have GC take
place in between and it go for weak/soft references.

I used 1.5.0_04 and checked the source code for ThreadLocal is 1.33
which is the same version as in 1.5.0_06 and the issue was initially
found using 1.5.0_06.

The issue is for real.
You can blame ThreadLocal but it does what it says on the tin.

Regards

Andy


-----Original Message-----
From: Robert Engels [mailto:rengels@ix.netcom.com]
Sent: 23 March 2006 16:05
To: java-dev@lucene.apache.org; rengels@ix.netcom.com
Subject: RE: [jira] Updated: (LUCENE-529) TermInfosReader and other +
instance ThreadLocal => transient/odd memory leaks =>
OutOfMemoryException

The only other thing that may be causing your problem is the use of
finalize(). This can interfere with the GC ability to GC objects.

I am now sure why the finalize() is used in the Lucene ThreadLocal
handling. It doesn't seem necessary to me.

-----Original Message-----
From: Robert Engels [mailto:rengels@ix.netcom.com]
Sent: Thursday, March 23, 2006 10:00 AM
To: java-dev@lucene.apache.org
Subject: RE: [jira] Updated: (LUCENE-529) TermInfosReader and other +
instance ThreadLocal => transient/odd memory leaks =>
OutOfMemoryException
Importance: High


Your testcase is invalid.

Reduce the size by 10, increase the repeat by 10, (SAME amount of memory
use), and it works fine.

The reason it works in the one case is that you use new
WeakReference(new Arrary()), - since the array cannot be referenced, it
is immediately GC'd. You should have noticed since there were no
finalization messages printed. You can demonstrate this clearly by
adding an else to the if in the finalize() to print out that the object
was indeed finalized.

ThreadLocal's work and are GC'd correctly.

There is something else wrong in your system.

I ran the test using 1.5.0._06.


-----Original Message-----
From: Andy Hind (JIRA) [mailto:jira@apache.org]
Sent: Thursday, March 23, 2006 7:07 AM
To: java-dev@lucene.apache.org
Subject: [jira] Updated: (LUCENE-529) TermInfosReader and other +
instance ThreadLocal => transient/odd memory leaks =>
OutOfMemoryException


     [ http://issues.apache.org/jira/browse/LUCENE-529?page=all ]

Andy Hind updated LUCENE-529:
-----------------------------

    Attachment: ThreadLocalTest.java

Attached is a test which you can use to see how ThreadLocals are left
around.
Getting an out of memory exception depends on a number things....it is
set up to fail for 64M

Now I understand what is going on, there are a few alternatives:

1) set null on close
- fine for single thread use
- probably leaves (n-1)*segments*2    things hanging around for n
threaded use

2) Use a weak reference and leave it up to GC to get rid of the referent
when it is not being used

3) Manage the things youself by object id and thread id - and clean up
on object close()

I would go with option 1) and 2) although it may mean things get GCed
before a call to close() when not used.

The fix I initially suggested is in production, and has been stress
tested with a couple of hundred users continually pounding the app,
 but not for multithreaded use of IndexReaders. Each time does a couple
of simple searches with no clever reuse of index readers (which is on
the todo list)

I do not see how setting the thread local to null on close() has any
negative impact. You are not going to use the cached information again??

Before the fix: 10-100 threads - 1G JVM - OOM in a few hours
After: 10-100 threads 256M JVM -  days with a flat memory footprint

I am not sure why the thread local table is so big for us, but that is
not really the issue.
It could just be building lots of IndexReaders (with thread locals
hanging - probably making 10/instance ) and gc not kicking in so this
table grows and can hold a lot of stale entries.  I may get time to
investigate further

> TermInfosReader and other + instance ThreadLocal => transient/odd
memory leaks =>  OutOfMemoryException
>
------------------------------------------------------------------------
-------------------------------
>
>          Key: LUCENE-529
>          URL: http://issues.apache.org/jira/browse/LUCENE-529
>      Project: Lucene - Java
>         Type: Bug
>   Components: Index
>     Versions: 1.9
>  Environment: Lucene 1.4.3 with 1.5.0_04 JVM or newer......will aplpy
to 1.9 code
>     Reporter: Andy Hind
>  Attachments: ThreadLocalTest.java
>
> TermInfosReader uses an instance level ThreadLocal for enumerators.
> This is a transient/odd memory leak in lucene 1.4.3-1.9 and applies to
current JVMs,
> not just an old JVM issue as described in the finalizer of the 1.9
code.
> There is also an instance level thread local in SegmentReader....which
will have the same issue.
> There may be other uses which also need to be fixed.
> I don't understand the intended use for these variables.....however
> Each ThreadLocal has its own hashcode used for look up, see the
ThreadLocal source code. Each instance of TermInfosReader will be
creating an instance of the thread local. All this does is create an
instance variable on each thread when it accesses the thread local.
Setting it to null in the finaliser will set it to null on one thread,
the finalizer thread, where it has never been created.  There is no
point to this :-(
> I assume there is a good concurrency reason why an instance variable
can not be used...
> I have not used multi-threaded searching, but I have used a lot of
threads each making searchers and searching.
> 1.4.3 has a clear memory leak caused by this thread local. This use
case above is definitely solved by setting the thread local to null in
the close(). This at least has a chance of being on the correct thread
:-)
> I know reusing Searchers would help but that is my choice and I will
get to that later ....
> Now you wnat to know why....
> Thread locals are stored in a table of entries. Each entry is *weak
reference* to the key (Here the TermInfosReader instance)  and a *simple
reference* to the thread local value. When the instance is GCed its key
becomes null.
> This is now a stale entry in the table.
> Stale entries are cleared up in an ad hoc way and until they are
cleared up the value will not be garbage collected.
> Until the instance is GCed it is a valid key and its presence may
cause the table to expand.
> See the ThreadLocal code.
> So if you have lots of threads, all creating thread locals rapidly,
you can get each thread holding a large table of thread locals which all
contain many stale entries and preventing some objects from being
garbage collected.
> The limited GC of the thread local table is not enough to save you
from running out of memory.
> Summary:
> ========
> - remove finalizer()
> - set the thread local to null in close()
>   - values will be available for gc

--
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators:
   http://issues.apache.org/jira/secure/Administrators.jspa
-
For more information on JIRA, see:
   http://www.atlassian.com/software/jira


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org


Mime
View raw message