[
https://issues.apache.org/jira/browse/AXIS2-4749?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12881292#action_12881292
]
Stefano Bruna commented on AXIS2-4749:
--------------------------------------
Yes, I also thought that using only the hash could lead to some false positive
duplicate messages as i wrote in the very last line in the issue description.
What we faced was that the 5 minutes task was too slow to clean the memory
under heavy load and the VM was reporting us many OutOfMemory.
Is in this situation the entire ESB node is not working until the cleaning
thread does its job. This is why we turned off the at-most-once interceptor as
you suggest.
In the case the interceptor was needed I was thinking about a FIFO queue base
on actule size of the received messages. When you reach a limit, older messages
are cleaned even if they are still within the 5 minutes time window.
I tryed to implement it but calculate the actual size of received messages is
uselessly complex and slow while here perfomance are really needed.
I'll reduce the 5 minutes timeout and increase server memory :)
> Tribe's AtMostOnceInterceptor could lead to OutOfMemory under heavy load and
> big messages.
> ------------------------------------------------------------------------------------------
>
> Key: AXIS2-4749
> URL: https://issues.apache.org/jira/browse/AXIS2-4749
> Project: Axis2
> Issue Type: Bug
> Affects Versions: 1.6
> Environment: Linux 64 bit, JDK 1.6, WSO2 Carbon 2.0.3
> Reporter: Stefano Bruna
>
> Wa are stressing the WSO2 ESB with some cache mediator enabled. Messages are
> echanged by tribes through the cluster's nodes. Under heavy load and with
> some big xml messages (1 mb per message) the local variable
> Map<ChannelMessage, Long> receivedMessages is growing continuously leading to
> a potential Out of Memory if the cleaning thread that runs every 5 minutes is
> not fast enough to free up memory.
> See:
> http://svn.apache.org/viewvc/axis/axis2/java/core/trunk/modules/clustering/src/org/apache/axis2/clustering/tribes/AtMostOnceInterceptor.java?view=markup
> This is because when an object is passed to a HashMap as a key internally the
> original object is kept. The Hasmap is using a private class Entry that has
> the hash, the orginal object used to create the hash, and the value that, in
> this usage case, is a long.
> Somewhere inside HashMap.java
> // private variables
> final Object key;
> Object value;
> Entry next;
> final int hash;
> // constructor
> Entry(int i, Object obj, Object obj1, Entry entry)
> {
> value = obj1; <==== value, ok
> next = entry; <===== entry with the same hash
> key = obj; <======= orginal object, needed for the equals use in
> case of same hash
> hash = i; <======== key, ok
> }
> So if the AtMostOnceInterceptor manages for example 50 msg/sec and a message
> is 1 mb we could have within 5 minutes a memory usage for the messageReceived
> object of 6 GB.
> A simple solution, if we dont'to accept all this as something by desing,
> could be to pass to the HashMap the already calculated hash of the object
> (that is also the same method that is called internally int i =
> hash(obj.hashCode());) to not give the opportunity to the HashMap to keep the
> actual object used to produce the key.
> .....
> // map
> private static final Map<Integer, Long> receivedMessagesHashCodes =
> new HashMap<Integer, Long>();
> .....
> Integer hashCode = new Integer(msg.hashCode());
>
> if (receivedMessagesHashCodes.get(hashCode) == null) { // If it is a new
> message, keep track of it
> receivedMessagesHashCodes.put(hashCode, System.currentTimeMillis());
> super.messageReceived(msg);
> } else { // If it is a duplicate message, discard it. i.e. dont call
> super.messageReceived
> log.info("Duplicate message received from " +
> TribesUtil.getName(msg.getAddress()));
> }
> etc...
> but maybe is not strong enough in the case tow messages have the same hash. A
> FIFO queue with a limited capacity ?
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
