river-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Peter Jones <...@roundroom.net>
Subject Re: DGC threads issue
Date Fri, 13 Jan 2012 14:33:12 GMT
Which is why the Object.equals/hashCode for an endpoint implementation is critical.

-- Peter


On Jan 13, 2012, at 2:05 AM, Gregg Wonderly wrote:

> I would say, that it's very easy to just code up a configuration entry, or a dynamic
construction in code where a new endpoint is also created per each Exporter.  That can quickly
turn into a problematic situation in cases like this, where there are lots of "quick" exports
followed by termination without "unexport" being done directly as part of the returning context.
> 
> Gregg Wonderly
> 
> On Jan 13, 2012, at 12:01 AM, Peter Firmstone wrote:
> 
>> Is there another way to create an Endpoint per exported object?  I'm just thinking,
it seems unlikely that Brian's implemented his own Endpoint, but are there any other error
conditions or incorrect use scenarios that could produce the same problem?
>> 
>> Cheers,
>> 
>> Peter.
>> 
>> Peter Jones wrote:
>>> Bryan,
>>> 
>>> DGC threads should not be per exported object.  Generally speaking, they tend
to be per endpoint (at which there are one or more remote objects exported).  Are you using
any sort of custom endpoint implementation?  Problems like this can occur when an endpoint
implementation doesn't implement Object.equals and hashCode appropriately, so the expected
batching of threads (and communication) per endpoint does not occur.
>>> 
>>> It might help to see, from a thread dump, exactly which DGC threads are causing
this problem.  And they are in the server JVM (with all the exported remote objects), not
the remote callers' JVM(s)?
>>> 
>>> -- Peter
>>> 
>>> 
>>> On Jan 12, 2012, at 3:45 PM, Tom Hobbs wrote:
>>> 
>>> 
>>>> Hi Bryan,
>>>> 
>>>> Sorry that no one got back to you about this.  I'm afraid that I don't
>>>> know the answer to your question, I've copied the dev list into this
>>>> email in case someone who monitors that list (but not this one) has
>>>> any ideas.
>>>> 
>>>> Best regards,
>>>> 
>>>> Tom
>>>> 
>>>> On Thu, Jan 12, 2012 at 2:29 PM, Bryan Thompson <bryan@systap.com>
wrote:
>>>> 
>>>>> Just to follow up on this thread myself.  I modified the pattern to return
a "thick" future rather than a proxy for the future.  This caused the RMI call to wait on
the server until the future was done and then sent back the outcome.  This "fixed" the DGC
memory/thread leak by reducing the number of exported proxies drammatically.
>>>>> 
>>>>> In terms of best practices, is distributed DGC simply not useful for
exported objects with short life spans?  Can it only be used with proxies for relatively long
lived services?
>>>>> 
>>>>> Thanks,
>>>>> Bryan
>>>>> 
>>>>> 
>>>>>> -----Original Message-----
>>>>>> From: Bryan Thompson
>>>>>> Sent: Tuesday, January 03, 2012 12:06 PM
>>>>>> To: user@river.apache.org
>>>>>> Subject: DGC threads issue
>>>>>> 
>>>>>> Hello,
>>>>>> 
>>>>>> Background:
>>>>>> 
>>>>>> I am seeing what would appear to be one DGC thread allocated
>>>>>> per exported object.  This is using River 2.2 and Sun JDK
>>>>>> 1.6.0_17.  Relevant configuration parameters are below.
>>>>>> 
>>>>>> I am observing problems with the DGC threads not being
>>>>>> retired on a timely basis.  The exported objects are proxies
>>>>>> for Futures which are being executed on the service.  The
>>>>>> code pattern is such that the proxied Future goes out of
>>>>>> lexical scope quite quickly.  E.g.,
>>>>>> rmiCallReturningProxyForFuture().get().
>>>>>> 
>>>>>> Under a modest load, a large number of such Futures are
>>>>>> exported which results in a large number of long lived DGC
>>>>>> threads.  This turns into a problem for the JVM due to the
>>>>>> stack allocation per thread.  Presumably this is not good for
>>>>>> other reasons as well (e.g., scheduling).
>>>>>> 
>>>>>> I have tried to override the leaseValue and checkInterval
>>>>>> defaults per the configuration options below.  I suspect that
>>>>>> the lease interval is somehow not being obeyed, which is
>>>>>> presumably a problem on my end.  However, I can verify that
>>>>>> the configuration values are in fact showing up in
>>>>>> System.getProperties() for at least some of the JVMs involved
>>>>>> (the one which drives the workload and the one that I am
>>>>>> monitoring with the large number of DGC lease threads).
>>>>>> 
>>>>>> Some questions:
>>>>>> 
>>>>>> Is this one-thread-per-exported proxy the expected behavior
>>>>>> when DGC is requested for the exported object?
>>>>>> 
>>>>>> The DGC lease checker threads appear to expire ~14 - 15
>>>>>> minutes after I terminate the process which was originating
>>>>>> the RMI requests.  This is close the sum of the default
>>>>>> leaseValue (10m) and checkInterval (5m) parameters, but maybe
>>>>>> there is some other timeout which is controlling this?  If
>>>>>> this is the sum of those parameters, why would the DGC lease
>>>>>> threads live until the sum of those values?  I thought that
>>>>>> the lease would expire after the leaseValue (10m default).
>>>>>> 
>>>>>> Can the issue I am observing be caused by a low heap pressure
>>>>>> on the JVM to which the RMI proxies were exported?  If it
>>>>>> fails to GC those proxies, even though they are reachable,
>>>>>> could that cause DGC to continue to retain those proxies on
>>>>>> the JVM which exported them?
>>>>>> 
>>>>>> Is there any way to configure DGC to use a thread pool or to
>>>>>> have the leases managed by a single thread?
>>>>>> 
>>>>>> Is it possible that there is an interaction with the useNIO option?
>>>>>> 
>>>>>> Relevant options that I am using include:
>>>>>> 
>>>>>>  -Dcom.sun.jini.jeri.tcp.useNIO=true
>>>>>>  -Djava.rmi.dgc.leaseValue=30000
>>>>>>  -Dsun.rmi.dgc.checkInterval=15000
>>>>>>  -Dsun.rmi.transport.tcp.connectionPool=true
>>>>>> 
>>>>>> Thanks in advance,
>>>>>> Bryan
>>>>>> 
>>> 
>>> 
>>> 
>> 
> 


Mime
View raw message