[jboss-jira] [JBoss JIRA] (JGRP-2057) Message bundler improvements

[Bela Ban (JIRA)]

    [ https://issues.jboss.org/browse/JGRP-2057?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13197340#comment-13197340 ] 

Bela Ban commented on JGRP-2057:
--------------------------------

The longer it takes to add a message to the hashmap, the more messages are sent by the bundler and that increases _throughput_. Apparently, lock contention on the send-queue in {{TransferQueueBundler}} between the senders (producers) and the sending thread (consumer) leads to more messages accumulating in the queue and therefore bigger message bundles with more messages.

{{SenderSendsBundler}} has much less contention, only an AtomicLong is incremented and decremented, and so fewer messages are in a batch, with the result being less throughput. 

Latency, however, is slightly better with {{SenderSendsBundler}}. Plus, its performance gets much better with more sending threads.

> Message bundler improvements
> ----------------------------
>
>                 Key: JGRP-2057
>                 URL: https://issues.jboss.org/browse/JGRP-2057
>             Project: JGroups
>          Issue Type: Enhancement
>            Reporter: Bela Ban
>            Assignee: Bela Ban
>             Fix For: 4.0
>
>
> * {{SingletonAddress}} is not needed anymore as we removed the shared transport in 4.0
> ** This slows down the bundler as it is used as key to the bundler hashmaps
> * {{SenderSendsBundler}}: send messages outside the lock scope. This requires us to use a number of bundler hashmaps, or making a shallow copy of the buckets of the current hashmap, before clearing it



--
This message was sent by Atlassian JIRA
(v6.4.11#64026)