Address-based Route Reflection Ruichuan Chen (MPI-SWS) Aman Shaikh (AT&T Labs - Research) Jia Wang (AT&T Labs - Research) Paul Francis (MPI-SWS) CoNEXT.

Address-based Route Reflection

Ruichuan Chen (MPI-SWS)Aman Shaikh (AT&T Labs - Research)Jia Wang (AT&T Labs - Research)Paul Francis (MPI-SWS)

CoNEXT 2011

2

iBGP reality Original full-mesh iBGP

All routers in an AS peer with all others Scales poorly

Route reflection Introduces routing anomalies

AS confederations

3

MED-based oscillation (MED: Multi-Exit Discriminator)

Routing anomalies in RR

C1

RR1

C3

RR2

C2

20

AS2AS1

2 1 3

1

10 30

y

IGP metric = x

MED = y

x

MEDs aren’t comparable between routes learned from different ASes!

RR1 prefers C2 > C1 > C3 > C2 > …

BGP best path decision:……Step 4: Lowest MED……Step 6: Lowest IGP metric……

4

Routing anomalies in RR Topology-based oscillation

C1

RR1

5

RR2

RR3

C2 C3

55

2

22

x IGP metric = x

Each RR is closer to another RR’s client.

5

Routing anomalies in RR

Forwarding loop

Path inefficiency

6

Routing anomalies in RR Why?

Different routers learn different routes!

Common solution RRs always prefer the routes learned from

clients over those from non-clients. Restricts RR placement. Not sufficient!

Design

Full-mesh’s semantics&

RR’s scalability

7

8

Our contribution Address-Based Route Reflection (ABRR)

Is decentralized Solves all oscillations and forwarding loops Has no path inefficiency Puts no restriction on RR placement Operates with no new BGP message

formats

Key insight BGP best-path decision for any given

prefix at a given router is highly dependent on information provided by other routers.

BGP best-path decision for any given prefix is independent of that of any other prefix.

9

10

Comparison Topology-Based Route Reflection (TBRR) Address-Based Route Reflection (ABRR)

Topology-Based Route Reflection Address-Based Route Reflection

11

ABRR - Client behaviors Each client peers with all ARRs

Advertise best route (if not iBGP-learned) to associated ARR

Advertise best route to eBGP neighbors

C1

ARR1

C3

ARR2

C2

AP: 0.0.0.0/1 AP: 128.0.0.0/1

Prefix: 1.2.0.0/16

AP: short for Address Partition

Prefix: 200.0.0.0/8

ABRR - RR behaviors

R2

TRR1 R1

R4

R3TRR2

Full-mesh: router learns multiple routes per prefixTBRR: router learns one single route per prefix

Full-mesh vs. TBRR Full-mesh TBRR

12

ABRR - RR behaviors (cont.)

Semantics equal to Full-mesh? ARR advertises multiple routes Routes flow through only one ARR

Scalability comparable to TBRR? Address partition

13

14

ABRR - RR behaviors (cont.) Each ARR peers with all clients

Advertise all best AS-level routes to clients ARR redundancy

Multiple ARRs for each AP

C1

ARR1

C3

ARR2

C2

AP: 0.0.0.0/1 AP: 128.0.0.0/1

1. Highest Local Preference2. Shortest AS Path3. Lowest Origin Type4. Lowest MED5. eBGP-learned over iBGP-learned6. Lowest IGP Metric7. Lowest Router ID8. Lowest Peer Address

AS-level criteria

15

No routing anomalies

ABRR emulates full-mesh semantics

No path inefficiency Full-mesh emulation Placement of ARRs within ISP network is

irrelevant

16

No routing anomalies (cont.)

No MED-based oscillation

No topology-based oscillation

No forwarding loop

Performance analysis

17

18

RIB-In and RIB-Out sizes of RRs

Parameters # routers: 2000 # APs / clusters: 50 # ARRs / TRRs per AP / cluster: 2 # peer ASes: 30

# best AS-level routes per prefix

RIB-In and RIB-Out sizes of RRs

For all parameter settings, ABRR has substantially smaller RIB size!

#APs or #Clusters #APs or #Clusters

19

20

iBGP sessions of RRs Modern routers or general-purpose

computers can handle thousands of sessions each with the full routing table.

iBGP convergence time

Implementation results

21

Fully functional implementation based on Quagga Less than 2K lines of code

Tier-1 ISP trace over 2 weeks

22

23

RIB-In and RIB-Out sizes

TBRR

24

# updates during two weeks

TBRR

25

Conclusion Address-Based Route Reflection

Solves all oscillations Finds efficient paths Places no constraints on RR placement Operates with no new BGP message

formats

Semantics equal to full-mesh Scalability comparable to TBRR