|
|
|
Jon Crowcroft |
|
University of Cambridge in London… |
|
|
|
|
|
|
|
IP routing currently is hierarchical |
|
ISPs are “tiered” with core, transit and access
ISPs |
|
Intra-as, typically Dijkstra SPF (OSPF or ISIS)
single destination, opportunistic (“greedy”), non pinned & symmetric |
|
Inter-AS is path vector policy constrained,
single destination, greedy, non pinned and asymmetric |
|
|
|
|
|
|
IntraAS multipath exists (c.f OSPF-SMP), mainly
for load balancing traffic on disaster recovery paths (rather than leaving
them idle/”hot standby”) |
|
InterAS multipath does not currently exist in
the routing infrastructure, although |
|
However: Multi-homing does, at both layers |
|
|
|
|
Intra-AS exists, sort of QOSPF |
|
Its quite hard – (NP hard for more than additive
+ 1 non additive parameter |
|
throughput and jitter are all non-additive metrics and loss
and delay may be |
|
Inter-AS we have zilch (except hacks based on
multi-homing) |
|
In theory could do shortest widest or widest
shortest, but user still needs to see ful MESH of BGP path data! |
|
|
|
|
|
|
Creative incentive to rationalise inter-AS
routes by allowing users choice |
|
Just as with congestion pricing, its not pricing
congestion, it’s the threat of pricing that creates pressure/evolutionary
trend – similarly, threat of user route alternates creases same |
|
Requires routing system to export alternatives
so user can exercise choice |
|
|
|
|
|
Policy = Router Information Filter |
|
Ingress, transit or egress |
|
Applied to who we advertise to, and |
|
Who we accept advertisements from |
|
Unilateral |
|
Followed by Path Vector Computation |
|
There are some global fields (AS registration,
and community attributes) |
|
|
|
|
|
|
|
Can use global attr and create a “long haul
provider selection” overlay |
|
Implement 3 ways: |
|
Loose source route |
|
Dynamic Address allocation (draw address from
space of desired) next hop provider |
|
Address Translation (whether IPv4 NAT, or IPv6
GSE like) |
|
|
|
|
|
|
Need to offer users possibility of path choice |
|
Ingress selection not sufficient |
|
May (just may) also need information brokering
about conditions |
|
Problem of thrashing of address/route tables if
we use hacks (already known/seen in BGP _ c.f. Huston et al) |
|
Problem of traffic engineering thrashing if user
dynamics are bad… |
|
Research needed… |
|
Notes
