[jira] Commented: (AMQ-2567) Zero Prefetch not working

["Rudolf Janz (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/activemq/browse/AMQ-2567?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=56908#action_56908 ] 

Rudolf Janz commented on AMQ-2567:
----------------------------------

Gary Tully wrote:
PrefetchSubscription and prefetchExtension on a deliveryAck in a transaction 
are areas that you need to look into to resolve this. Great that you have a 
simple junit test case.
---

You are right, prefetchExtension is the problem. I do not understand the reason for it. If I want prefetch 0, it should stay that way. Right now it is increased on each pull.

In acknowledge there is the following comment 
      if (context.isInTransaction()) {
              // extend prefetch window only if not a pulling consumer
                                if (getPrefetchSize() != 0) {
                                    prefetchExtension = Math.max(
                                            prefetchExtension, index );

In PrefetchSubscription::pullMessage it checks for prefetchSize 0 
        if (getPrefetchSize() == 0 && !isSlave()) {
            final long dispatchCounterBeforePull;
        	synchronized(this) {
        		prefetchExtension++;
        		dispatchCounterBeforePull = dispatchCounter;
        	}


> Zero Prefetch not working
> -------------------------
>
>                 Key: AMQ-2567
>                 URL: https://issues.apache.org/activemq/browse/AMQ-2567
>             Project: ActiveMQ
>          Issue Type: Bug
>          Components: Broker
>    Affects Versions: 5.3.0
>         Environment: Unix/Windows, Java 1.6
>            Reporter: Rudolf Janz
>         Attachments: ZeroPrefetchConsumerTest.java
>
>
> I have problems using the zero prefetch. The consumer starts prefetching messages after some receives. 
>  
> Our use case is, that we have a large computation which is split into smaller sub jobs. These jobs are sent via ActiveMQ to some processing nodes on different machines. The duration of jobs differs very much (10s to some minutes). The jobs are sent by decreasing estimated computation time. If one of the consumers which receives a large job prefetches some other jobs, these will be processed later. In the meantime the remaining consumers are idle, and the total computation time is much longer than necessary.
>  
> I have modified the existing ZeroPrefetchConsumerTest to test for the problem (I have removed the other test methods). 
>  
> Two consumers (C1  and C2) are instantiated. These messages are sent:
> 1,2,3,4,5,6,7,8,9
>  
> C1 reads 4 times, receives 1,2,3,4 -> this is correct
>  
> now C2 reads, it receives 8 not 5, which is the next message in the queue. The reason is, that C1 prefetched 5, 6, 7, that should not have happened.(sometimes C1 only prefetches 5,6)
>  
> The problem can be seen in the JMX Console as well, after a while, the first consumer has more than one dispatched message and the queue has an InflightCount of 3, although there are only two consumers!
>  
> The last version that we used was 4.1.1, that worked.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.