Route Control Platform Making an AS look and act like a router

Route Control Platform – IEEE CCW 2004

1

Route Control PlatformMaking an AS look and act like a router

Aman Shaikh

AT&T Labs - Research

IEEE CCW 2004

Matt Caesar (UC Berkeley)Don Caldwell (AT&T Labs – Research)

Nick Feamster (MIT)Jennifer Rexford (AT&T Labs – Research)

Kobus van der Merwe (AT&T Labs – Research)


2

You have heard the news• BGP is broken

– It converges slowly– At times it does not converge at all– It causes routing loops and deflections inside an AS– It’s misconfigured frequently– Traffic engineering is hard with BGP

• Fixing BGP is hard– Incremental fixes

• Makes BGP even more complicated

– New architectures and inter-domain protocols• Deployment is almost impossible


3

What are the fundamental problems?• AS is a logical entity for inter-domain routing

(i.e. BGP) and yet BGP state and logic are decomposed across routers inside an AS– No router has complete BGP state– Each router makes routing decision based on partial

and incomplete view

• BGP interacts in odd ways with other protocols– Most notably with the IGP (Interior Gateway

Protocol) running inside an AS


4

Fixing the fundamental problems:“Route Control Platform”

• Represents an AS as a single logical entity– Complete view of AS’s routes– Computes routes for all routers inside an AS

• Routers no longer have to compute routes

• Exchanges routing information with RCPs in other ASes

AS 3AS 2AS 1

iBGP

Physicalpeering

Inter-AS ProtocolRCP RCP RCP


5

The rest of this talk: the case for RCP• Principles for inter-domain routing

– Treat the AS as a whole and compute routes using AS-wide state• Example: high-level policy expression

– Control routing protocol interactions• Example: interaction between BGP and IGP

• Potential dealbreakers– Backwards compatibility and incentives

– Scalability and reliability

• Related work (or…haven’t we seen this before?)– Route reflection and route servers

– Overlay networks


6

Decomposed configuration state

Sprint UUNet

A C10.0.0.1 192.168.0.1

Simple policy: “Don’t advertise routes learned from UUNet to Sprint”Configuration is decomposed, so routes must carry state

neighbor 192.168.0.1 route-map IMPORT-C inroute-map IMPORT-C permit 10 set community 0:1000

ip community-list 1 permit 0:1000neighbor 10.0.0.1 route-map EXPORT-A outroute-map EXPORT-A deny 10 match community 1

Assign routes “From UUNet” tag at router C

Don’t send route with “From UUNet” tag to Sprint at router A


7

RCP: Centralize configuration

• RCP implements policies for entire AS– Knows about sessions to all other ASes– Implements policies in terms of relationship with

ASes

• Benefits– Simpler configuration– Do not have to tag routes with state

Sprint UUNet

A C10.0.0.1 192.168.0.1

RCP


8

BGP interacts with underlying protocols

RR1 RR2

C1 C2

3 3

1

1 1

d

C1 learns BGP route to destination from RR1C2 learns BGP route to destination from RR2

C1 sends packets to RR1 via its IGP shortest path which traverses C2

C2 sends packets to RR2 via its IGP shortest path which traverses C1

Persistent forwarding loop


9

RCP: Compute routes with complete info

• RCP learns all externally learned routes• Computes consistent router-level paths• Benefits:

– Intrinsic loop freedom and convergence– RCP does not have to stick to BGP decision process

• Can “pin” paths for traffic engineering and other purposes

RR1 RR2

C1 C23 3

1

1 1

d

“RR2”“RR2” RCP


10

Getting from here to there: three easy steps

• Two key issues– Backward compatibility

– Deployment incentives

AS 3AS 2AS 1

iBGP

Physicalpeering


iBGP

eBGP


11

Phase 1: Control over Protocol Interactions

iBGP

eBGP

Before: conventional iBGP

iBGP

eBGP

After: RCP gets “best” iBGP routes (and IGP topology)

Only one AS has to change its architecture!

RCP


12

Phase 1 Application: Controlling Path Changes

BGP routes take “nearest exist” (shortest IGP path)Failures or maintenance can change IGP (path) weights

Exit point can also changeTraffic shifts, convergence delay, congestion

A B

C D


13

Phase 1 Application: Controlling Path Changes

BGP routes take “nearest exist” (shortest IGP path)Failures or maintenance can change IGP (path) weights

RCP can “pin” exit points as IGP weights change

RCPA B

C D


14

Phase 2: AS-wide Selection and Policy

iBGP

eBGP

Before: RCP gets “best” iBGP routes (and IGP topology)

After: RCP gets all eBGP routes from neighbors

iBGP

eBGP

RCP

RCP


15

Phase 2 Application: Efficient Aggregation

192.168.0.0/23192.168.0.0/24

192.168.0.0/23192.168.1.0/24

192.168.0.0/23 (??) 192.168.0.0/23 (??)iBGP

eBGP

Aggregation curbs routing table growth

Routers can’t know which routers need more specific routes


16

Phase 2 Application: Efficient Aggregation

192.168.0.0/23192.168.0.0/24

192.168.0.0/23192.168.1.0/24

192.168.0.0/23 192.168.0.0/23

Aggregation curbs routing table growth

192.168.0.0/24 192.168.1.0/24

192.168.0.0/23192.168.0.0/23

RCP

RCP can determine which routers need more specific routesand which routers can do away with less specific routes


17

Phase 3: All ASes have RCPsBefore: RCP gets all eBGP routes from neighbors

iBGP

eBGP

After: ASes exchange routes via RCP

RCP

AS 3AS 2AS 1

iBGP

Physicalpeering



18

Phase 3 Application: More Flexible Routing

• Better network management– Diagnostics and trouble-shooting– Routing co-located with other information (e.g.

traffic)– Ability to reason about an AS as a single entity

• Protocol Improvements– Attaching prices to routes– Inter-AS negotiation of exit points– Overlay routing informed by IP-layer information

• Your application here…


19

Scalability and Robustness• Can RCP scale?

– We have a prototype implementation• Single-box RCP can handle AS-wide BGP load• OSPF changes can be troublesome

– Centralized != unable to scale

• Is RCP a single point of failure?– RCP can be implemented using distributed system

insights– Consistency (mostly) a non-isssue

• Guarantee from OSPF/IS-IS operation:– “Either an RCP replica has a complete view of network

(partition) or no view; but never a partial view”


20

Is RCP basically a Route Reflector?• Yes, but it’s a better route reflector

• “Customized” routing decisions for clients– Route reflectors do not compute routes from client’s

perspective– Route reflectors do not emulate a “full mesh”

• Routing decisions based on complete visibility– Guaranteed correct routes– Replication is dictated by system issues


21

RCP also looks a lot like…• A “route server”

– Route arbiter: looked at applying policy at exchange points

– AS agents• RCP can act as an AS agent; can answer queries for the AS

• An overlay network– Most previous work is in data overlays– RCP is a control overlay

• Hierarchical routing is about control overlays

– RCP could give more information and control to data overlays• RCP has AS-wide information and direct control over paths

taken through the AS


22

Conclusions• RCP embodies two principles for inter-domain

routing– Treat an AS as a single logical entity

• Compute consistent routes using complete AS-wide view

– Control routing protocol interactions

• Benefits– Simpler, more expressive configuration– Intrinsic robustness: no loops, faster convergence– Enable new applications and innovations

• Opportunity for new traffic engineering applications


23

More informationFDNA 04 paper“The Case for Separating Routing from Routers”Nick Feamster, Hari Balakrishnan, Jennifer Rexford, Aman Shaikh,

Kobus van der Merwe

http://www.research.att.com/~ashaikh/publications.html

Documents

Route Control Platform Making an AS look and act like a router