[Oisf-devel] Better PF_RING performance with workers run-mode now instead of autofp

[Victor Julien]

On 01/11/2012 07:21 PM, Martin Holste wrote:
> Thanks for the info Chris--our workload is very similar to yours, so
> it's good to have that confirmation.
> 
> Update: ac/autofp is better than ac/auto, with something like 20%
> drop, and 800% CPU compared with about 500% using workers.  So, I
> think I'm going back to ac-gfbs/workers unless anyone else has a tweak
> to suggest.
>
> My hunch is that ac/auto performs poorly because only one thread is
> assigned to do the stream and decode, which at high volumes is intense
> and too much for a single thread to handle, especially now that server
> responses have the HTTP buffers added.

Yes, "auto" doesn't scale very well. Autofp should be better if you have
more than 4 cores I think. We're working on an improved load balancer
for the autofp runmodes btw. Initial results look promising.

-- 
---------------------------------------------
Victor Julien
http://www.inliniac.net/
PGP: http://www.inliniac.net/victorjulien.asc
---------------------------------------------