RoundRobinConnectionDistributor useDedicatedAcceptor

classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

RoundRobinConnectionDistributor useDedicatedAcceptor

Daniel Feist
Hi,

Just a very quick question on this.  Was wondering if there was any
data, or if any tests had been performed on the impact of using (or
not) a dedicated acceptor.  I know jetty's model is to use dedicated
acceptors(s), but I'm unsure of the value of doing this.

One potential disadvantage with this option that occured to me is on
hardware with large number of cores, when using a selector per core,
you may have a situations where the ratio of acceptors vs selectors is
low.

4 core   - 1 acceptor, 3 selectors
8 core   - 1 acceptor, 7 selectors
24 core - 1 acceptor, 23 selectors.

thanks,
Dan
Reply | Threaded
Open this post in threaded view
|

Re: RoundRobinConnectionDistributor useDedicatedAcceptor

oleksiys
Administrator
Hi,

I think using dedicated acceptor is common practice in production
environment.
However for some cases, when you don't expect connection to
connect/disconnect much - you can share acceptor to not waste a thread.
So effectively you'll have:

4 core - 4 selectors
8 core - 8 selectors
24 core  - 24 selectors

Thanks.

WBR,
Alexey.

On 09.12.14 03:46, Daniel Feist wrote:

> Hi,
>
> Just a very quick question on this.  Was wondering if there was any
> data, or if any tests had been performed on the impact of using (or
> not) a dedicated acceptor.  I know jetty's model is to use dedicated
> acceptors(s), but I'm unsure of the value of doing this.
>
> One potential disadvantage with this option that occured to me is on
> hardware with large number of cores, when using a selector per core,
> you may have a situations where the ratio of acceptors vs selectors is
> low.
>
> 4 core   - 1 acceptor, 3 selectors
> 8 core   - 1 acceptor, 7 selectors
> 24 core - 1 acceptor, 23 selectors.
>
> thanks,
> Dan