[Bridge] SMP vs UNP performance

[Bridge] SMP vs UNP performance

[Joubert Berger]

I was running a frame loss test the other day against an rhel4 smp
kernel and a non-smp kernel.  I noticed I was getting better numbers
when I ran the non-smp kernel.  I was surprised by that.  Anyone have
any insight into this?  I am running this with e1000 dirvers.

I also ran a test where I ran the smp kernel with hyper threading
turned on and off.  There, the frame loss numbers were worse when
hyper threading was turned off.  I would not have thought that having
hyperthreading turned off would make such a big difference.

Just trying to make sense of these numbers :-)

--joubert

Re: [Bridge] SMP vs UNP performance

[Stephen Hemminger]

On Sun, 27 Aug 2006 10:05:01 -0400
"Joubert Berger" <joubertb@gmail.com> wrote:

> I was running a frame loss test the other day against an rhel4 smp
> kernel and a non-smp kernel.  I noticed I was getting better numbers
> when I ran the non-smp kernel.  I was surprised by that.  Anyone have
> any insight into this?  I am running this with e1000 dirvers.

Turning NAPI on/off might change the numbers slightly.

> I also ran a test where I ran the smp kernel with hyper threading
> turned on and off.  There, the frame loss numbers were worse when
> hyper threading was turned off.  I would not have thought that having
> hyperthreading turned off would make such a big difference.
> 
> Just trying to make sense of these numbers :-)
> 
> --joubert

Forwarding packets is inherently single threaded, the additional CPU's
can't really make it any faster. They can only cause additional cache
misses and locking overhead. The additional CPU's might help if you
performed a bidirectional or multiple interface test. Forwarding packets
shouldn't normally CPU bound, it is limited by the bus and interrupt rate.
If you are doing lots of firewalling that can change.

Re: [Bridge] SMP vs UNP performance

[Joubert Berger]

Thanks Stephen,

I will go back and see what happens happens in a bi-directional test.

And yes, we plan on using the firewall capabilities, so I will spend
some time looking at what happens when I am running with all the
firewall rules loaded.

--joubert

On 8/28/06, Stephen Hemminger <shemminger@osdl.org> wrote:
> On Sun, 27 Aug 2006 10:05:01 -0400
> "Joubert Berger" <joubertb@gmail.com> wrote:
>
> > I was running a frame loss test the other day against an rhel4 smp
> > kernel and a non-smp kernel.  I noticed I was getting better numbers
> > when I ran the non-smp kernel.  I was surprised by that.  Anyone have
> > any insight into this?  I am running this with e1000 dirvers.
>
> Turning NAPI on/off might change the numbers slightly.
>
> > I also ran a test where I ran the smp kernel with hyper threading
> > turned on and off.  There, the frame loss numbers were worse when
> > hyper threading was turned off.  I would not have thought that having
> > hyperthreading turned off would make such a big difference.
> >
> > Just trying to make sense of these numbers :-)
> >
> > --joubert
>
> Forwarding packets is inherently single threaded, the additional CPU's
> can't really make it any faster. They can only cause additional cache
> misses and locking overhead. The additional CPU's might help if you
> performed a bidirectional or multiple interface test. Forwarding packets
> shouldn't normally CPU bound, it is limited by the bus and interrupt rate.
> If you are doing lots of firewalling that can change.
>