"Packet Shapers"

Sat Aug 1 19:33:51 UTC 1998

> > i'm not stating that the value is only at the edge.  i'm stating that i
> > don't think it is realistic to be sticking transparent devices that hold
> > detailed state information on flows in the core.
>
> Inktomi's CTO has publicly stated that this is their objective, although
> the phrase used was "super aggregation nodes" and not "core". If you're
> interested, ask him for his 3W3 paper submission wrt handling 100 Gbs
> links.

I have seen Paul Gauthier's paper.  Last time I spoke with Paul & Vince (VP
Field Ops) when I was out in the bay area, that is precisely what we
discussed, among other things.  Sure, they can theoretically scale to those
types of links.

Their suggestion is to use policy based routing (PBR) instead of Alteon (or
similiar L2/L4 technology) and the reliance of an Ethernet being avaiable
for drop in of their technology.  Anyone who knows anything about PBR knows
that PBR is neither fault tolerant nor scalable nor something I want to be
force to use for this type of stuff.  Anyone who knows anything about
technologies like WCCP knows that it is more than just plain PBR, because
state is maintained in the L3 switching engine.  The Alteon does maintain
state of its cache farm memebers, but not of the sessions themselves.

Other major issues (aside from their problematic network deployment
strategy) that I see around Inktomi are in the current platform choice.  It
is neither cost effective nor does it make sense from a performance
standpoint.

I have discussed these with numerous people at Inktomi over the past months
and not gotten a solution.

> > there is enormous value in the core: some people get it right,
> some get it
> > horribly wrong.  but lets face it - the core of a networks
> primary function
> > is to move packets fast, and move them well.
>
> Violently agree with this.

Ditto.  You need to do much more than just push packets fast.

You need to be able to distinguish services on the network, perhaps on a per
flow basis, and make sure your mission critical services or high end
customers don't get trampled on.  You need to be able to push packets fast
_and_ do better than just plain best effort.  You need to be able to do this
fault tolerant and/or highly available.

Cheers,
Chris

--
Christian Kuhtz, BellSouth Corp., Sr. Network Architect  <ck at bellsouth.net>
1100 Ashwood Parkway, Atlanta, GA 30338                        <ck at gnu.org>