2008.02.19 NANOG 42 Lightning talk #5 network costs in the core

2008.02.19 Lightning talk #5

One more round tomorrow, submit your
ideas now!

Curtis Villamizar, Drew Perkins

Total network costs in the core.
glimpse at key findings; look at slides
Made up topology, came up with core nodes,
core links onto it; presented to non-IP-aware
group at Stanford.

66 cities modeled, 18 core cities, you have
small city-to-city traffic flows, you aggregate
them into large core-to-core traffic flows.
costs with various interfaces.

costs for various network interfaces.

Router costs for OC768 on large routers;
aimed for demands comperable to smaller ISPs
on left, and larger ISPs are on the right,
and then scaled out to right into future.

At low utilization, you have wasted, unused
capacity on OC768; for 10gig, similar curve,
but much; $700/Mb for OC, $400/Mb for 10g

costs for l2/l3 10gigE; smaller capacity,
you need to build a CLOS, an array of routers
to handle capacity, you get beyond one; but
router costs go WAY down, so your transport
costs dominate, rather than router costs,
and drops you to $250/Mb

UUnet did IP router bypass
parallel 10gig links between sets of cities;
if you put links bypassing routers, you can
eliminate interface card costs, save money.

Practical limit to IP bypass, based on the
traffic matrix needs and failover requirements.
at top end, bypass ends up being full mesh,
which is also impractical.

IP router bypass link cost savings can be very
substantial, if you can use them; but the
practicality is limited by the number you can
effectively make use of.

Even with free transport, full mesh is not
cost effective.

Impact of MPLS TE...lots of interesting directions
they're going into.
OC768 tops out over OC192, which tops 10gigE