Reduced State Routing in the Internet

1
Reduced State Routing in the Internet

• Ramakrishna Gummadi
• Collaborators Nupur Kothari, Young-Jin Kim,
  Ramesh Govindan, Brad Karp, and Scott Shenker

2
Motivation

• Router economics
• Forwarding state expensive
• Routing state cheap
• Current routing protocols equate these, and the
  state in routers scales with the size of the
  Internet (modulo aggregation)
• Basic question can we design a routing
  architecture where core routers keep forwarding
  state proportional to their fanout?

3
Inspiration Geographic Routing in Wireless
Networks

• Key property routing state proportional to node
  density
• Addressing geographic locations
• Common-case greedy routing
• Exception perimeter routing
• Right-hand face traversal on planarized sub-graph

4
Geographic Routing for the Internet?

• Superficially attractive, but next hop choice
  constrained purely by geography
• Does not allow us to express
• Inter-domain BGP policies
• Intra-domain Traffic Engineering policies

5
Reduced State Routing (RSR)

• Clean-slate re-design of Internet node addressing
  and packet forwarding
• No IP addresses!
• Relatively small changes to routing protocols
• Properties
• Compact forwarding state at core routers
• No major change in forwarding state at edge
• No major change in routing state

6
RSR Addressing and Routing

• Addressing nodes identified by geographic
  location
• Routing current two-level routing hierarchy
• Inter-domain routing
• Uses BGP with minor modifications
• Intra-domain routing
• Greedy routing on a high-dimensional embedding of
  the ISP's traffic engineering requirements
• Similar to GNP, different motivation
• Fall back to geographic routing when greedy
  routing fails

7
Intra-domain Routing

• ISP represents traffic engineering constraints as
  IGP link weights

• Each router computes shortest path matrix from
  IGP link-state database

• Each router computes a high-dimensional embedding
  (using the textbook Bourgain algorithm) of this
  matrix
• Distance between two routers is approximately the
  shortest path distance

8
Forwarding Plane Example
A
2.7
1
A
A
D
D
D
2
2
1.8
2
2
2
2
C
2.9
C
B
C
B
3
3
B
9
Forwarding Plane Example
A
2.7
1
A
A
D
D
D
2
2
1.8
2
2
2
2
C
C
B
C
B
2.9
3
3
B
10
Inter-domain Routing

• BGP used to compute best egress edge router from
  every ingress edge router in AS1 to AS2
• Edge router ER1 creates a forwarding table entry
  with key AS2 and value egress edge router ER2

11
Inter-domain Forwarding
lookup
AS2
D
write
ER2
AS2
D

• Source writes destination AS number into packet
• ER1 looks up and writes embedding address of ER2
  into packet

12
Transformation Into Intra-domain Forwarding
ER2
ER1
AS1

• ER1 starts Intra-domain forwarding with
  high-dimension embedding of ER2 as destination

13
Evaluation Questions

• How big are packet headers?
• How well does RSR perform compared to shortest
  path routing?
• What is the forwarding state at core routers?

14
Packet Header

• Larger packet header than today
• Need six address entries
• How big do the virtual coordinates need to be?
• How many dimensions? What resolution per
  dimension?
• For the ISPs we examined, 16 dimensions and 5
  bits per dimension work well

15
Stretch

• Path stretch ratio of total weight along a path
  taken by RSR to the least cost path
• Average stretch average over all pair-wise path
  stretches
• 2 for all ISP graphs with completely
  un-optimized embedding
• Two components
• Greedy on VCs efficient (lt1.1)
• Geographic routing killer (50)
• Two hop neighboring state to reduce geographic
  routing impact

16
Core Router State

• Small (lt40) number of forwarding entries for ISP
  topologies
• Purely a function of average node degree
  (fanout), and independent of number of total
  nodes or edges

17
Summary and Future Work

• Exhibited a reduced state routing architecture
  for the Internet
• Radically different forwarding paradigm
• Small changes to routing protocols packet
  contents
• Preserves routing policy and control
• Naïve implementation with reasonable stretch
• Future directions
• Improve embedding quality
• Are link weights sufficient to capture traffic
  engineering?
• Alternates to geographic routing?