[
https://issues.apache.org/jira/browse/AMQ-5920?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14659838#comment-14659838
]
ASF subversion and git services commented on AMQ-5920:
------------------------------------------------------
Commit 340728f2d1eb4c1aa74f4002c5b16bf7a70c57b7 in activemq's branch
refs/heads/master from [~gtully]
[ https://git-wip-us.apache.org/repos/asf?p=activemq.git;h=340728f ]
AMQ-5920 - use implicit broker transaction for virtual topic fanout and add
concurrentSend=true option to use an executor to fanout. Combination gives a 3x
reduction in fanout roundtrip for small persistent messages to 100 consumer
queues
> Improve performance of virtual topic fanout
> -------------------------------------------
>
> Key: AMQ-5920
> URL: https://issues.apache.org/jira/browse/AMQ-5920
> Project: ActiveMQ
> Issue Type: Improvement
> Components: Broker
> Affects Versions: 5.11.0
> Reporter: Gary Tully
> Assignee: Gary Tully
> Labels: kahadb, perfomance, virtualTopic
> Fix For: 5.12.0
>
>
> Virtual topics provide a nice alternative to durable subs. Each durable sub
> is modeled as a separate queue.
> There are performance implications however, because a message has to be sent
> to each of the (fanout) queues. For a durable subs, there is a single message
> in the journal and just index updates for each sub.
> To improve performance there are three ways to improve the comparison between
> virtual topics and durable subs.
> # avoid the disk sync associated with enqueue
> # do parallel enqueues to get the store batching writes
> # introduce message references in the journal to reduce the disk io
> For 1, introducing a transaction (either client side or automatically, broker
> side) ensures there is a single disk sync on commit.
> For 2, using an executor to do sends in parallel allows the journal to batch
> as seen in AMQ-5077
> For 3, the implementation is a lot more involved; for recovery there needs to
> be a journal write per destination and reading the journal will require two
> reads because of the indirection. Tracking gc needs to be handled to ensure
> the referenced entry is maintained. In short this is a lot of work and will
> only be visible for large (> 8k) messages where the cost of a large v small
> journal write is noticeable. The messageId dataLocator provides an entry
> point for this work but considering that 1 & 2 combined can give a 3x
> improvement I don't think it is worth the effort (and added complexity) at
> the moment.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
