On AS-Level Path Inference

Jia Wang (AT&T Labs Research) Joint work withZ. Morley Mao (University of Michigan, Ann Arbor)Lili Qiu (University of Texas, Austin) Yin Zhang (University of Texas, Austin)

On AS-Level Path Inference 2June 10, 2005

Discover end-to-end forwarding path between two hosts

Berkeley

Internet

CNN

Calren

Level3

GNN

Qwest Sprint UUnet

University company

AT&T

business

ISP ISP ISP

ISP ISP ISP ISP

ISP

Berkeley

Calren

Level3 Qwest Sprint UUnet

University company

AT&T

business

ISP ISP ISP

ISP ISP ISP ISP

ISP

Berkeley

Calren

Level3 Qwest Sprint UUnet

University company

AT&T

business

ISP ISP ISP

ISP ISP ISP ISP

ISP

On AS-Level Path Inference 3June 10, 2005

Motivation

Network diagnoses Performance optimization Overlay network Content distribution Network modeling

On AS-Level Path Inference 4June 10, 2005

Example – overlay routing

SourceDestination

?

?? ?

154

2

3

Internet

On AS-Level Path Inference 5June 10, 2005

Outline

Related work Routescope Evaluation Improvements

AS relationship inferenceFirst AS hop inference

Conclusion

On AS-Level Path Inference 6June 10, 2005

Related work

Forwarding path discoveryWith direct access to the source

Router-level: traceroute AS-level: [Mao Sigcomm2003] [Mao Infocom2004]

Without direct access to the source None!

On AS-Level Path Inference 7June 10, 2005

Challenges

Asymmetric routing Over 60% of AS paths asymmetric

Complicated routing policies Not shortest path routing Commercial relationship between ASes determines

how traffic flow though the Internet

Multi-homing Very common

On AS-Level Path Inference 8June 10, 2005

Routescope

Key observation: relationships among ASes play important role in determining feasible forwarding paths

Approach: Infer AS-level paths by finding the shortest policy path in an AS graph obtained from BGP tables collected from multiple vantage points

On AS-Level Path Inference 9June 10, 2005

Assumptions

Explicit AS relationships Peer-peer Provider-customer

Shortest AS policy path preferred “Valley-free” rule

Uniform routing policy within an AS AS destination based uniform routing Stability

These assumptions are mostly correct.

On AS-Level Path Inference 10June 10, 2005

AS relationships translate into BGP export rules Export to a provider or a peer

Allowed: its routes and routes of its customers and siblings

Disallowed: routes learned from other providers or peers

Export to a customer or a siblingAllowed: its routes, the routes of its customers

and siblings, and routes learned from its providers and peers

On AS-Level Path Inference 11June 10, 2005

“Valley-free” rule

After traversing a provider-customer or peer-peer edge, cannot traverse a customer-provider or peer-peer edge

Invalid path: >= 2 peer links downhill-uphilldownhill-peerpeer-uphill

On AS-Level Path Inference 12June 10, 2005

Example of valley-free paths

XX

[1 2 3], [1 2 6 3] are valley-free

[1 4 3], [1 4 5 3] are not valley free

On AS-Level Path Inference 13June 10, 2005

AS path inference algorithm

Compose the AS graph based on BGP tables Infer AS relationship Classify edges based on AS relationship

Customer-provider (UP) link Provider-customer (DOWN) link Peer-peer (FLAT) link

Compute shortest policy path conforming the “valley-free” rule using modified Dijkstra’s algorithm

Infer the first AS hop if multiple paths returned

On AS-Level Path Inference 14June 10, 2005

Evaluation

Based on existing AS relationship inference algorithms Gao: based on the degree of ASes along the path SARK: consider AS hierarchy properties BPP: formulate as 2SAT problem and develop

heuristics that yield minimum of invalid paths Compare AS-level paths

Extracted from a large number of BGP tables Among 125 public BGP gateways

On AS-Level Path Inference 15June 10, 2005

Paths in BGP tables

BGP table

AlgorithmUnique paths

MatchMatch length

Exact match

Shorter Longer

AS7018 (tier-1)

Gao

18085

77% 80% 33% 18% 2%

SARK 67% 79% 34% 15% 4%

BPP 84% 85% 37% 15% 0%

AS2152 (tier-2)

Gao

11990

62% 65% 10% 34% 1%

SARK 48% 57% 29% 40% 3%

BPP 67% 67% 12% 33% 0%

AS8121 (tier-3)

Gao

15757

16% 27% 3% 69% 4%

SARK 14% 23% 3% 72% 4%

BPP 18% 30% 3% 66% 5%

On AS-Level Path Inference 16June 10, 2005

Paths between BGP gateways

BGP gateway

AlgorithmUnique paths

MatchMatch length

Exact match

Shorter Longer

All

Gao

2457

30% 51% 21% 15% 35%

SARK 38% 61% 24% 20% 19%

BPP 18% 29% 15% 5% 66%

US

Gao

1907

24% 43% 16% 18% 40%

SARK 40% 57% 24% 24% 19%

BPP 22% 42% 18% 10% 48%

BPP yields most accurate AS path inference than GAO and SARK

On AS-Level Path Inference 17June 10, 2005

Possible causes of mismatches

Inaccuracy in AS relationship inferenceEspecially in non-North American regions

Multihoming

On AS-Level Path Inference 18June 10, 2005

Inaccurate AS relationship inference 19%~66% of inferred paths are longer than

actual paths Significant inconsistency among AS relationship

inference results

Common peer-peer Common provider-customer

Gao vs SARK 229 (3.63%, 36.12%) 41730 (89.43%, 94.68%)

Gao vs BPP 5959 (94.51%, 48.42%) 39606 (84.87%, 97.74%)

SARK vs BPP 334 (52.68%, 2.71%) 33752 (85.66%, 93.17%)

Solution: infer more accurate AS relationships

On AS-Level Path Inference 19June 10, 2005

A new AS relationship inferencealgorithm Problem formulation: integer programming

Each edge e in the direct graph G = (V,E) Relation(e) = 1 (customer-provider), 2 (peer-peer), or 3 (provider-

customer) Constraints

If r is reverse edge of e, relation(e)+relation(r) = 4. Every path in use is valley-free, i.e., for (e1,e2) on a path, relation(e1)

= 1 relation(e2) = 3. For any (src,dst), if there is a path P that is shorter than actual path,

then P is not valley free, i.e., (e1,e2) on P s.t. relation(e1) ≠ 1 relation(e1) ≠ 3.

Novelty: derive additional constraints that violate valley free constraints

Solution: improved random walk algorithm [Selman et al. 1993] Handle non-binary variables Repeatedly remove stub ASes with out-degree of 0

On AS-Level Path Inference 20June 10, 2005

AS path inference with accurate AS relationship

Unique paths

MatchMatch length

Exact match

Shorter Longer

AS7018 (tier-1) 18085 82% 83% 35% 17% 0%

AS2152 (tier-2) 11990 64% 64% 10% 35% 0%

AS8121 (tier-3) 15757 16% 27% 3% 69% 4%

All BGP gateways 2457 70% 73% 30% 22% 4%

US BGP gateways 1907 60% 62% 27% 34% 4%

The accuracy is among the best of other three in BGP table experiments and is much higher than alternatives in BGP gateway experiments.

On AS-Level Path Inference 21June 10, 2005

Multihoming

Over half of the mismatches occur at the very first hop AS

If first hop is known, over 15% of mismatches can be eliminated

Solution: infer the first hop AS

AS S

AS T2

AS T1 AS D

AS C

Inferred path

Actual path

On AS-Level Path Inference 22June 10, 2005

First hop inference

Gather candidate first hop ASes from S by launch traceroute to S from multiple vantage points

Identify the transition point T that is likely to be on the path from S to D by testinghop_count(S,T) + hop_count(T,D) = hop_count(S,D)

SourceDestination

AS S

AS T2

AS T1 AS D

AS C

Transition point T1

T2 Assume having access to D

On AS-Level Path Inference 23June 10, 2005

Hop count inference

Hop_count(S,T) ≈ hop_count(T,S) Hop_count(H,D): H = S or T

Send ping packet to H Guess the initial TTL value TTL0 set by H Get TTL value TTL1 in ICMP response packet received from H Hop_count(H,D) = TTL0 - TTL1 + 1

Common value for TTL0 32 (Win95/98/Me) 64 (Linux, Compaq Tru64) 128 (Win NT/2000/XP) 255 (most UNIX systems)

On AS-Level Path Inference 24June 10, 2005

Improvement with known first AS hop

Unique paths Match length Improvement

AS7018 (tier-1) 18085 86% 3%

AS2152 (tier-2) 11990 76% 12%

AS8121 (tier-3) 15757 48% 21%

All BGP gateways 1907 70% 8%

US BGP gateways 2457 88% 15%

On AS-Level Path Inference 25June 10, 2005

Possible causes of inaccuracy

Complicated AS relationships: 15% paths Two consecutive FLAT links DOWN link followed by a FLAT link FLAT link followed by UP link Dual transit/peering relationship

Routing policies Shortest path vs. customer routes Inconsistent advertisement to different peering locations BGP tie-breaking rules

AS prepending > 28% ASes

On AS-Level Path Inference 26June 10, 2005

Conclusion

Routescope: AS-level path inference tool without access to the source

Two enhancements AS relationship inference First hop inference

Accuracy: up to 88% inferred paths have the same length as the actual paths

New metric for evaluating AS relationship inference Evaluate existing AS relationship inference algorithms