Re: 3.0 wishlist Was: Overview of 2.2.x goals?

linux kernel account (
Mon, 19 Jan 1998 20:39:06 -0500 (EST)

Why not reconfigure the cards as the same enet address and do the load
balencing at the outgoing que (incoming would of course not need load

On Mon, 19 Jan 1998, John Carol Langford wrote:

> >The question is, would this incur significant overhead at high speeds?
> >Wouldnt just multipath+no-route-cache be better at e.g. 10mbps?
> Depends on the cpu. For a rough rule of thumb, ipeql on a Ppro-200
> multipathing between 2 fast ethernet cards would just saturate the cpu.
> One saturated fast ethernet card would consume about 30%-35% of the cpu.
> If ipeql incurred no overhead (currently due to an extra layer of indirection
> and not using the route cache), you would expect 60-70% cpu utilization.
> Most of the extra 30-40% overhead is (I expect) due to losing the route cache.
> The extra overhead may be acceptable on a fast cpu using only 10Mbps
> connections.
> The "best" solution would be to not waste the route cache - something which
> doesn't seem possible just in the realm of a device driver. Alterations to
> the IP stack changing the way hardware header caching is done would be
> necessary.
> -John Langford