[
https://issues.apache.org/jira/browse/ARTEMIS-994?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15883112#comment-15883112
]
Michael Andre Pearce (IG) commented on ARTEMIS-994:
---------------------------------------------------
[[email protected]] Exactly, we are running our Artemis brokers to the point
now we are down to 0.100 -> 0.200 millisecond end to end latency for non
persisted (non durable), and 0.200 -> 0.400 millisecond latency for persisted
(durable).
So hats off to you guys for such a low impact broker already, and why I'm
trying to get my company to adopt it :).
Once Clebert's fix for libaio, was applied, where we've had to be running NIO
to get this performance due to the SSD's we use, since switching to the new
aligned libaio version we saw cpu lower, and more importantly less jitter in
the latency, and also less gc's work.
What we aim to achieve here is the same as that does, decrease the cpu usage,
decrease the amount of gc, and also reduce the jitter due to these.
As Clebert states disk and io we cannot change, but what we can reduce is the
cpu work, and gc which contribute to jitters, in turn latency and throughput of
the system.
> Support Netty Native Epoll on Linux
> -----------------------------------
>
> Key: ARTEMIS-994
> URL: https://issues.apache.org/jira/browse/ARTEMIS-994
> Project: ActiveMQ Artemis
> Issue Type: New Feature
> Components: Broker
> Affects Versions: 2.0.0, 1.5.3
> Reporter: Michael Andre Pearce (IG)
>
> Netty has support for native epoll, this is available only on linux systems.
> This is more performant.
> http://netty.io/wiki/native-transports.html
> This is inline with supporting asyncio with libaio using such native feature
> to bring perfomance benefits to artemis.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
