Re: big picture UDP/IP performance question re 2.6.18 -> 2.6.32

[Willy Tarreau <w@1wt.eu>]

On Sat, Oct 01, 2011 at 02:16:05PM -0400, starlight@binnacle.cx wrote:
> At 08:44 AM 10/1/2011 +0200, Eric Dumazet wrote:
> >In my experience, I have the exact opposite :
> >performance greatly improved in recent
> >kernels. Unless you compile your kernel to include
> >new features that might reduce performance
> >(namespaces, cgroup, ...)
> 
> RH has both of the above turned on in the
> 2.6.32-71.29.1.el6.x86_64 kernel tested.
> 
> If these are big negatives to network
> performance, could you list what should
> specifically turned off to maximize
> results?  Also a recommendation for
> the best recent kernel for another
> benchmark would be helpful.
> 
> Probably can't convince anyone to deploy a
> kernel without commercial support, but if
> an alternate compile fixes performance it
> might be possible to convince RH to support
> the alternative build.

Just a suggestion, instead of measuring CPU usage at a
given load, could you check what maximal load you can
achieve ? It is very possible that CPU usage report is
not accurate. We observed this in a number of situations,
especially in high packet rate environments where the
usage is a sum of many micro-measurements.

Also, I did not notice any indication on the load level
you were reaching (packets per second and bandwidth).
Have you compared the interrupt rate ? It is possible
that they differ between the two kernels, for instance
because the NIC auto-adapts instead of being throttled
to a given rate. This can have a significant impact on
measurements and performance.

Regards,
Willy