ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Valentin Kulichenko <valentin.kuliche...@gmail.com>
Subject Re: Cleaning internal thread locals
Date Wed, 06 Apr 2016 19:16:29 GMT
Guys,

Thread locals that are not cleaned up is a bug in our code, because this is
a memory leak. This started with Tomcat only because it has automatic
detection of such leaks and informs about them. Other application servers
can easily lack this feature.

Alexei, can you create a unit test first and list all the problematic
places in the code? I believe there will be no generic solution, we will
have to resolve all different cases individually.

-Val

On Wed, Apr 6, 2016 at 2:51 AM, Denis Magda <dmagda@gridgain.com> wrote:

> Cleaning of thread local variables like the ones below should be avoided
> otherwise the performance will be impacted
>
> OptimizedObjectStreamRegistry$StreamHolder
> BinaryMemoryAllocatorChunk
>
> I wouldn't clean these variable even from tomcat filter when a Thread from
> Tomcat's pool finished processing a request because this Thread can be used
> to interact with Ignite later.
>
> Simple full-text search shows that Ignite uses ThreadLocals in tons of the
> places. So from my understanding it doesn't make sense to investigate every
> variable and decide how to deal with this.
>
> Alexei, is there any way to get from Tomcat its internal threads on
> context destroy and clean all our ThreadLocals? If to clean variables from
> the filter then we have to do it wisely avoiding cleaning of performance
> related ThreadLocals.
> If we use renewThreadsWhenStoppingContext will these messages disappear?
>
> --
> Denis
>
>
> On 4/6/2016 12:30 PM, Semyon Boikov wrote:
>
>> As I understand in many palces thread locals can be used for performance
>> reasons, and it is not good idea to always clear it.
>>
>> Initial problem in IGNITE-967 was about tomcat, since tomcat provides
>> solution for this problem do we really need fix something in Ignite?
>>
>> Valentin, you created this ticket, what do you think?
>>
>> On Wed, Apr 6, 2016 at 11:59 AM, Alexei Scherbakov <
>> alexey.scherbakoff@gmail.com> wrote:
>>
>> It is a very good idea to try to fix internal thread local cleaning in all
>>> places.
>>> On example, it can be easily done in class GridSpinReadWriteLock
>>> Just replace code in line 172 with
>>> readLockEntryCnt.remove();
>>> and add line 388 with:
>>> if (update == 0) readLockEntryCnt.remove();
>>> ( needs more testing possibly )
>>>
>>> I've provided you with the list of offending places ( which may not be
>>> completed )
>>> But it is definitely a good place to start.
>>> Who will be responsible for fixiing it ?
>>>
>>>
>>> 2016-04-06 11:50 GMT+03:00 Denis Magda <dmagda@gridgain.com>:
>>>
>>> As far as I understand Alexei, the issue here is that Ignite kernal
>>>> doesn't have list of all the threads that accessed it during its
>>>>
>>> lifetime.
>>>
>>>> So the kernal won't be able to iterate over all the threads (like
>>>> threads
>>>> from Tomcat thread pool) and perform all the cleaning. This is the
>>>> issue.
>>>>
>>>> --
>>>> Denis
>>>>
>>>>
>>>> On 4/6/2016 11:47 AM, Artem Shutak wrote:
>>>>
>>>> Alexei,
>>>>>
>>>>> I would follow the Valentin's suggestion in the issue description and
>>>>> would
>>>>> add to Ignite tests a check on thread locals after stopping all grids.
>>>>>
>>>> And
>>>
>>>> then we will be able to fix thread local cleaning in all places.
>>>>>
>>>>> Thanks,
>>>>> -- Artem --
>>>>>
>>>>> On Wed, Apr 6, 2016 at 11:18 AM, Alexei Scherbakov <
>>>>> alexey.scherbakoff@gmail.com> wrote:
>>>>>
>>>>> Valentin,
>>>>>
>>>>>> I did test on demo web app where HttpSession is backed by Ignite
cache
>>>>>> and
>>>>>> collect some logs.
>>>>>> There are many places in Ignite where thread locals are used.
>>>>>> Then cache methods are accessed in threads from servlet container
>>>>>> pool,
>>>>>> we
>>>>>> get leaks.
>>>>>> I don't think it will be easy to add thread locals cleaning in all
>>>>>>
>>>>> these
>>>
>>>> places without major code rewrite, but, for sure, it would be ideal
>>>>>> solution.
>>>>>>
>>>>>>
>>>>>> <E0><EF><F0> 05, 2016 4:21:52 PM
>>>>>> org.apache.catalina.loader.WebappClassLoaderBase
>>>>>> checkThreadLocalMapForLeaks
>>>>>> SEVERE: The web application [/app1] created a ThreadLocal with key
of
>>>>>> type
>>>>>> [org.apache.ignite.internal.util.GridSpinReadWriteLock$1] (value
>>>>>> [org.apache.ignite.internal.util.GridSpinReadWriteLock$1@5fecc35f])
>>>>>>
>>>>> and
>>>
>>>> a
>>>>>> value of type [java.lang.Integer] (value [0]) but failed to remove
it
>>>>>> when
>>>>>> the web application was stopped. Threads are going to be renewed
over
>>>>>> time
>>>>>> to try and avoid a probable memory leak.
>>>>>> <E0><EF><F0> 05, 2016 4:21:52 PM
>>>>>> org.apache.catalina.loader.WebappClassLoaderBase
>>>>>> checkThreadLocalMapForLeaks
>>>>>> SEVERE: The web application [/app1] created a ThreadLocal with key
of
>>>>>> type
>>>>>> [org.apache.ignite.internal.util.tostring.GridToStringBuilder$1]
>>>>>> (value
>>>>>>
>>>>>>
>>> [org.apache.ignite.internal.util.tostring.GridToStringBuilder$1@4ecdfc04
>>>
>>>> ])
>>>>>> and a value of type [java.util.LinkedList] (value
>>>>>>
>>>>>>
>>>>>>
>>> [[org.apache.ignite.internal.util.tostring.GridToStringThreadLocal@684a2791
>>>
>>>> ]])
>>>>>> but failed to remove it when the web application was stopped. Threads
>>>>>>
>>>>> are
>>>
>>>> going to be renewed over time to try and avoid a probable memory leak.
>>>>>> <E0><EF><F0> 05, 2016 4:21:52 PM
>>>>>> org.apache.catalina.loader.WebappClassLoaderBase
>>>>>> checkThreadLocalMapForLeaks
>>>>>> SEVERE: The web application [/app1] created a ThreadLocal with key
of
>>>>>> type
>>>>>> [org.apache.ignite.internal.binary.BinaryThreadLocalContext$1] (value
>>>>>> [org.apache.ignite.internal.binary.BinaryThreadLocalContext$1@4e9300e7
>>>>>>
>>>>> ])
>>>
>>>> and a value of type
>>>>>> [org.apache.ignite.internal.binary.BinaryThreadLocalContext] (value
>>>>>> [org.apache.ignite.internal.binary.BinaryThreadLocalContext@7934a88])
>>>>>> but
>>>>>> failed to remove it when the web application was stopped. Threads
are
>>>>>> going
>>>>>> to be renewed over time to try and avoid a probable memory leak.
>>>>>> <E0><EF><F0> 05, 2016 4:21:52 PM
>>>>>> org.apache.catalina.loader.WebappClassLoaderBase
>>>>>> checkThreadLocalMapForLeaks
>>>>>> SEVERE: The web application [/app1] created a ThreadLocal with key
of
>>>>>> type
>>>>>> [java.lang.ThreadLocal] (value [java.lang.ThreadLocal@56502b72])
and
>>>>>> a
>>>>>> value of type
>>>>>> [org.apache.ignite.internal.binary.streams.BinaryMemoryAllocatorChunk]
>>>>>> (value
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>> [org.apache.ignite.internal.binary.streams.BinaryMemoryAllocatorChunk@4eab9d2f
>>>
>>>> ])
>>>>>> but failed to remove it when the web application was stopped. Threads
>>>>>>
>>>>> are
>>>
>>>> going to be renewed over time to try and avoid a probable memory leak.
>>>>>> <E0><EF><F0> 05, 2016 4:21:52 PM
>>>>>> org.apache.catalina.loader.WebappClassLoaderBase
>>>>>> checkThreadLocalMapForLeaks
>>>>>> SEVERE: The web application [/app1] created a ThreadLocal with key
of
>>>>>> type
>>>>>> [java.lang.ThreadLocal] (value [java.lang.ThreadLocal@eb53c9d]) and
a
>>>>>> value
>>>>>> of type
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>> [org.apache.ignite.marshaller.optimized.OptimizedObjectStreamRegistry.StreamHolder]
>>>
>>>> (value
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>> [org.apache.ignite.marshaller.optimized.OptimizedObjectStreamRegistry$StreamHolder@5e2c15f7
>>>
>>>> ])
>>>>>> but failed to remove it when the web application was stopped. Threads
>>>>>>
>>>>> are
>>>
>>>> going to be renewed over time to try and avoid a probable memory leak.
>>>>>> <E0><EF><F0> 05, 2016 4:21:52 PM
>>>>>> org.apache.catalina.loader.WebappClassLoaderBase
>>>>>> checkThreadLocalMapForLeaks
>>>>>> SEVERE: The web application [/app1] created a ThreadLocal with key
of
>>>>>> type
>>>>>> [org.apache.ignite.internal.util.GridSpinReadWriteLock$1] (value
>>>>>> [org.apache.ignite.internal.util.GridSpinReadWriteLock$1@55b9da28])
>>>>>>
>>>>> and
>>>
>>>> a
>>>>>> value of type [java.lang.Integer] (value [0]) but failed to remove
it
>>>>>> when
>>>>>> the web application was stopped. Threads are going to be renewed
over
>>>>>> time
>>>>>> to try and avoid a probable memory leak.
>>>>>> <E0><EF><F0> 05, 2016 4:21:52 PM
>>>>>> org.apache.catalina.loader.WebappClassLoaderBase
>>>>>> checkThreadLocalMapForLeaks
>>>>>> SEVERE: The web application [/app1] created a ThreadLocal with key
of
>>>>>> type
>>>>>> [org.apache.ignite.internal.util.StripedCompositeReadWriteLock$1]
>>>>>>
>>>>> (value
>>> [org.apache.ignite.internal.util.StripedCompositeReadWriteLock$1@1ac73431
>>>
>>>> ])
>>>>>> and a value of type [java.lang.Integer] (value [9]) but failed to
>>>>>>
>>>>> remove
>>>
>>>> it
>>>>>> when the web application was stopped. Threads are going to be renewed
>>>>>> over
>>>>>> time to try and avoid a probable memory leak.
>>>>>> SEVERE: The web application [/app1] created a ThreadLocal with key
of
>>>>>> type
>>>>>> [org.apache.ignite.internal.processors.cache.GridCacheAdapter$5]
>>>>>> (value
>>>>>>
>>>>>>
>>> [org.apache.ignite.internal.processors.cache.GridCacheAdapter$5@28a1bd99
>>>
>>>> ])
>>>>>> and a value of type
>>>>>>
>>>>>>
>>>>>>
>>> [org.apache.ignite.internal.processors.cache.GridCacheAdapter.FutureHolder]
>>>
>>>> (value
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>> [org.apache.ignite.internal.processors.cache.GridCacheAdapter$FutureHolder@6763d70c
>>>
>>>> ])
>>>>>> but failed to remove it when the web application was stopped. Threads
>>>>>>
>>>>> are
>>>
>>>> going to be renewed over time to try and avoid a probable memory leak.
>>>>>>
>>>>>> As a solution, I propose to implement servlet filter as a part of
>>>>>>
>>>>> Ignite
>>>
>>>> web integration, which will be responsible for cleaning up all thread
>>>>>> locals after request processing.
>>>>>> Filter must be executed last in filter chain to work correctly.
>>>>>> It is not necessary to use the filter in case of Tomcat, because
it
>>>>>> provides own solution to the problem requiring special configuration
>>>>>> parameter for context
>>>>>> User should decide which approach is better to him.
>>>>>>
>>>>>>
>>>>>> 2016-04-06 1:45 GMT+03:00 Valentin Kulichenko <
>>>>>> valentin.kulichenko@gmail.com
>>>>>>
>>>>>> :
>>>>>>> Alexei,
>>>>>>>
>>>>>>> I would start with defining which thread locals cause this and
why
>>>>>>>
>>>>>> they
>>>
>>>> are
>>>>>>
>>>>>> not properly cleaned on node shutdown. Ideally, this should never
>>>>>>> happen.
>>>>>>>
>>>>>>> -Val
>>>>>>>
>>>>>>> On Tue, Apr 5, 2016 at 10:36 AM, Alexei Scherbakov <
>>>>>>> alexey.scherbakoff@gmail.com> wrote:
>>>>>>>
>>>>>>> Hello.
>>>>>>>
>>>>>>>> Currently I'm working on the IGNITE-967
>>>>>>>> <https://issues.apache.org/jira/browse/IGNITE-967>.
>>>>>>>>
>>>>>>>> In fact, it is well known, easily reproduceable, ClassLoader
memory
>>>>>>>>
>>>>>>>> leak
>>>>>>> problem in managed environments, such as servlet containers.
>>>>>>>
>>>>>>>> See http://wiki.apache.org/tomcat/MemoryLeakProtection to
read more
>>>>>>>>
>>>>>>> on
>>>
>>>> that.
>>>>>>>> In short, the code using thread locals is responsible for
cleaning
>>>>>>>>
>>>>>>> them
>>>
>>>> after request processing
>>>>>>>> to prevent memory leaks related to different lifetime of
application
>>>>>>>>
>>>>>>>> and
>>>>>>> thread in pool.
>>>>>>>
>>>>>>>> Also blindly reusing thread from pool can be dangerous, because
>>>>>>>>
>>>>>>> current
>>>
>>>> ThreadLocal value can be different from expected default value.
>>>>>>>>
>>>>>>>> One posiible way of solving the problem is using post-processing
>>>>>>>>
>>>>>>>> Filter,
>>>>>>> which is responsible for cleaning all thread locals on current
>>>>>>> thread.
>>>>>>>
>>>>>>>> Another approach would be sticking with container's specific
>>>>>>>> solution
>>>>>>>>
>>>>>>>> to
>>>>>>> this problem.
>>>>>>>
>>>>>>>> On example, tomcat provides the configuration property
>>>>>>>> renewThreadsWhenStoppingContext,
>>>>>>>> which forces it to renew threads involved in servicing requests
for
>>>>>>>> destroyed context, thus removing leaks.
>>>>>>>>
>>>>>>>> What do you think?
>>>>>>>>
>>>>>>>>
>>>>>>>> --
>>>>>>>>
>>>>>>>> Best regards,
>>>>>>>> Alexei Scherbakov
>>>>>>>>
>>>>>>>>
>>>>>>>> --
>>>>>>
>>>>>> Best regards,
>>>>>> Alexei Scherbakov
>>>>>>
>>>>>>
>>>>>>
>>> --
>>>
>>> Best regards,
>>> Alexei Scherbakov
>>>
>>>
>

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