Subject: Re: link0 and ifconfig
To: Lubomir Sedlacik <salo@Xtrmntr.org>
From: Mipam <mipam@ibb.net>
List: netbsd-users
Date: 08/09/2004 14:07:07
On Mon, 9 Aug 2004, Lubomir Sedlacik wrote:

> hi,
> 
> On Mon, Aug 09, 2004 at 11:49:37AM +0200, Mipam wrote:
> > 
> > > man ifconfig, search for link[0-2]:
> > 
> > Thanks for your answer.  I was actually searching for something else.
> > When you have a machine with need to digest high loads of network
> > traffic, the cpu is also loaded more to process the packets.  Checksum
> > offloading helps to save cpu cycles, but maybe a bigger bottleneck may
> > be the amount of interrupts that are being generated.  Normally for
> > each packet that is received, an interrupt is triggered.  I was
> > looking for IRQ mitigation to remedy this somewhat, so that these
> > interrupts are being surpressed until the nics FIFO is full. This
> > saves the amount of interrupts and thus saves cpu cycles.  Some cards
> > may support this feature, some intel cards do.  Is their support for
> > this feature?  Or are there other ways to reduce the amount of
> > interrupts.  FreeBSD also supports device polling to get ridd of
> > interrupt stuff, this works with some cards. I was wondering what i
> > can do reduce it?
> 
> fxp(4)
> 
>      Certain versions of the i8255x support loading microcode which
> implements a receive interrupt mitigation function, known as
> ``CPUSaver''.  Use of this option can improve performance in some
> situations by reducing interrupt load on the host.  This option is
> available on the following chip versions:
> 
>      o   i82558 step A4 (rev 4)
>      o   i82558 step B0 (rev 5)
>      o   i82559 step A0 (rev 8)
>      o   i82559S step A (rev 9)
>      o   i82550 (rev 12)
>      o   i82550 step C (rev 13)
> 
>      This option is enabled by setting the ``link0'' option with
> ifconfig(8).

Clear, thanks.
I did not see anything in the wm manpage, so this means that this featue 
is no longer present in the modern intel cards? Why not, because it didnt 
have too much effect anyway or? So basically, the amount of interrupts is not 
seen as a problem on machines which high loads off traffic?
So checksum offloading and later lso is much more effective than reducing 
the amount of interrupts?
Bye,

Mipam.