Thanks.
We had looked at the memory usage of 0-10 connections and seen a
similar effect when Robbie was fixing some other memory leaks. The
64K element Method array in the 0-10 Session objects is needed so
that sessions can synchronise with the client, although perhaps a
more efficient structure could be used to store them, such as a Map
that can grow and shrink dynamically? As you say, overall memory
usage is much reduced now, so the real issue is the failure to
deregister connections in the broker.
I will have to try your code from QPID-3162 to see if the same issue
still exists on the updated 0-10 network stack, since I remember
adding explicit calls to remove connections from the broker registry...
Looking at the patch, I have a few points:
* in MultiVersionProtocolEngine you set _delegate to null, same
with _protocolEngine in MINANetworkDriver, and the Mina IoSession
attachment is also set to null. Which of these was *actually*
preventing the GC from reclaiming memory? If we know that, then we
can remove only that reference. This many seems like overkill if the
session can be closed properly.
* in Connection, setState(CLOSED) is functionally equivalent to the
two lines you replace it with, since the monitor on lock is held
already.
* calling _connection.onOpen(null) from the closed method in
ProtocolEngine_0_10 seems odd. It sets the _onOpenTask of
ServerConnection to null, which you could do in setState(CLOSED) of
ServerConnection (see previous item) instead?
I will need to investigate a bit further, but adding a test case
seems like a useful thing to try and do.
Andrew.
--
-- andrew d kennedy ? do not fold, bend, spindle, or mutilate ;
-- http://grkvlt.blogspot.com/ ? edinburgh : +44 7582 293 255 ;
On 23 Mar 2011, at 19:06, Danushka Menikkumbura wrote:
Just run the sample client that I have mentioned in [1] and see the
number of ProtocolEngine_0_10 and related objects in heap on
JProfiler. You should be able to see 50 of them.
[1] - https://issues.apache.org/jira/browse/QPID-3162
Thanks,
Danushka
On Thu, Mar 24, 2011 at 12:13 AM, Danushka Menikkumbura
<[email protected]> wrote:
Hi Andrew,
I just checked the latest trunk. The issue is still there, the
memory consumption is very less, though. As I can see its because
of the way the incomplete Method array is allocated in Assembler
now. But still I can see the same result w.r.t. instances that are
not GC'ed.
Thanks,
Danushka
On Wed, Mar 23, 2011 at 6:07 PM, Andrew Kennedy
<[email protected]> wrote:
Also,
I will add a test case to the new network code. I haven't seen this
behaviour on the branch I'm working on, so it may be fixed. I'll
wait for your patch before checking again, though.
I should have more time to work on the in-vm networking branch now
that I've started a new job; I know I said it would be part of 0.10
but circumstances meant I didn't have enough time to finish the work.
Andrew.
--
-- andrew d kennedy ? do not fold, bend, spindle, or mutilate ;
-- http://grkvlt.blogspot.com/ ? edinburgh : +44 7582 293 255 ;
On 23 Mar 2011, at 10:47, Andrew Kennedy wrote:
Hi, Danushka.
I'll take a look at this, it looks like a race during session close,
which we have seen before.
Andrew.
--
-- andrew d kennedy ? edinburgh : +44 7582 293 255
On 23 March 2011 00:43, Danushka Menikkumbura
<[email protected]> wrote:
Hi,
Any idea what is going wrong here?.
Thanks,
Danushka
On Tue, Mar 22, 2011 at 12:30 PM, Danushka Menikkumbura <
[email protected]> wrote:
You can run a simple connection create/close loop and check the
number of
times MINANetworkDriver.sessionClosed gets called.
When I did that I observed this particular method got only called
49 time
when I ran the loop for 50 iterations, 99 times for 100 and 994 for
1000.
Am I missing something here?. I also see that the ProtocolEngine_0_10
instance count in heap is equal to the number of times this method
did not
get called. Hence the memory leak. These objects retain in the
memory as
they are not removed from the _typeMap in ConfigStore as
removeConfiguredObject does not get called.
Thanks,
Danushka
On Tue, Mar 22, 2011 at 5:09 AM, Danushka Menikkumbura <
[email protected]> wrote:
Hi devs,
I see sometimes (1/50) MINANetworkDriver.sessionClosed does not get
called
causing a memory leak. Is this a known issue?.
Thanks,
Danushka
---------------------------------------------------------------------
Apache Qpid - AMQP Messaging Implementation
Project: http://qpid.apache.org
Use/Interact: mailto:[email protected]
