Observations from Observations from Router-level TracesRouter-level Traces

Lisa AminiIBM T. J. Watson Research Center

Joint with Henning Schulzrinne, Aurel LazarColumbia University

ContextContext Near-term issues facing service providers

for Web-facing applications Mirrored servers

where to place servers? which network access providers? how many connection points? how many servers to deploy? where to direct clients?

End-to-end routing behavior/delay vs. connectivity

AS level focus

Beyond ConnectivityBeyond Connectivity Inter-domain (AS) routing tables

Aggregation Policy-based routing Oregon Route Server

Stub Networks

Transit Networks

A

CED

B

Autonomous Systems

TracerouteTraceroute Router-level end-to-end probing

Traceroute.org website AS lookup via Routing Arbiter Database (RADB) Round trip time (RTT) Loose source-routing Artifacts

D1 DatasetD1 Dataset 189 sites Random pairings

Forward and reverse paths Poisson arrivals Mean time between probes 10 minutes 220,551 measurements 5 days 1/3 US sites, 2/3 non-US (31 countries)

[ Zhang, Duffield, Paxson, Shenker]

Router-level Path LengthRouter-level Path Length mean=16.45, std. dev.=4.39

Num

ber

of R

oute

s

0 5 10 15 20 25 300

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Routes with US endpoints

Routes with European endpoints

All routes

Router Hops

AS Path LengthAS Path Length

0 5 10 15 20 25 300

0.5

1

1.5

2

2.5

3

3.5x 10

4

All routesUS only routesEuropean only routes

AS Hops

Num

ber

of R

oute

s mean=6.47, std. dev.=2.33

AS Degree FrequencyAS Degree Frequency Faloutsos3 -0.97 (or higher) correlation coefficient98% of nodes represented

AS degree Frequency

100

101

102

103

100

101

102

103

All Routes

Routes with US endpoints

Rank

0.97 correlation coefficient required discarding 31% of nodes

Which is correct?Which is correct?

Routing view traceroute target

traceroute source

AS FrequencyAS Frequency

AS frequency

100

101

102

103

100

101

102

103

104

All Routes

Routes with US endpoints

Rank

0.97 correlation coefficient required discarding 10% of nodes

Response Time PredictionResponse Time Prediction Does path length predict delay?

Probe Round Trip Time

Router Hops

mean RTT=263msCorrelation coefficient= 0.31

Correlation coefficient =0.27

AS Hops

AS PropertiesAS Properties Can we predict delay based on AS

path properties? Ranked each AS according to:

AS edge degree AS frequency

Calculated mean RTT per path length Grouped by top 5, 6-10, 10-20 AS’s

Influence of Path PropertiesInfluence of Path PropertiesHigh edge degree AS?

Number of Router Hops / Path

Average RTT

0 5 10 15 20 25 300

100

200

300

400

500

600Top 5 AS'sTop 10 AS'sTop 20 AS's

High Frequency AS?High Frequency AS?

0 5 10 15 20 25 300

100

200

300

400

500

600Top 5 AS'sTop 10 AS'sTop 20 AS'sAll AS's

Number of Router Hops / Path

Average RTT

High Edge Degree AS?High Edge Degree AS?High edge degree AS in 2 AS hops?

Number of Router Hops / Path

Average RTT

0 5 10 15 20 25 300

100

200

300

400

500

600Top 5 AS'sTop 10 AS'sTop 20 AS's

Backbone AS?Backbone AS?

Number of Router Hops / Path

Average RTT

0 5 10 15 20 25 300

100

200

300

400

500

600AS's ranked 1-5AS's ranked 6-10AS's ranked 10-20All routes

[http://ratings.metrix.net 7/2001]

Backbone AS in 2 AS hops?Backbone AS in 2 AS hops?

Number of Router Hops / Path

Average RTT

0 5 10 15 20 25 300

100

200

300

400

500

600AS's ranked 1-5AS's ranked 6-10AS's ranked 10-20Remaining paths

AS Affinity?AS Affinity?

Number of Router Hops / Path

Average RTT

0 5 10 15 20 25 300

100

200

300

400

500

60040% in an AS50% in an AS60% in an ASAll

Route AsymmetryRoute Asymmetry Paxson, 1995 Path(A,B) Path(B,A) Issues

Ping triangulation Congestion Avoidance Internet mapping

A B

AS Hop DifferencesAS Hop Differences

Number of AS hop differences

Number of Routes

0 2 4 6 8 10 12 1410

0

101

102

103

104

105

All RoutesUS Endpoint RoutesEU endpoint Routes

11755 paired, unique routes

57% routes were AS-path asymmetric

Compare with 30% based on 1995 data

74% asymmetric from first AS hop

Traceroute IssuesTraceroute Issues AS assignment

[Chang, Jamin, Willinger]

AS1

AS7AS6AS5

AS4AS3AS2

BGP AS PathBGP AS Path

Can we predict forward and reverse end-to-end metrics from BGP AS_PATH?

Looking Glass Probing 92 Sites 8372 unique path measurements 2202 fully paired (BGP forward and

reverse, traceroute forward and reverse)

AS Hop DifferencesAS Hop Differences

Number of AS hop differences

Number of Routes

0 5 10 15 200

500

1000

1500traceroute reverseBGP forwardBGP reverse

47% of forward paths correctly predicted by reverse traceroute

49% of forward paths correctly predicted by reverse BGP AS_PATH

69% of forward paths correctly predicted by forward BGP AS_PATH

34% asymmetry between forward/reverse BGP AS_PATH

SummarySummary BGP routing tables provide complete view

from single location. Aggregation Filtering

End-to-end probing from points throughout network can provide insights beyond connectivity Limited view of connectivity Traceroute issues with noise, node

assignment BGP AS_PATH inaccurate as path predictor

Discard CriteriaDiscard Criteria• Origin traceroute server not

responding• Incomplete traceroute output• Internal use only node address (e.g.,

10.x.x.x, 172.16.x.x-172.32.x.x ranges)• Route did note terminate in target AS• Intermediate node did not respond to

ICMP echo• No matching reverse probe for same

time period

What can we conclude?What can we conclude?

Results are for RTT only Negative results for

Path Length AS degree, AS frequency, Backbone,

hops from origin High delay as result of transient

conditions (congestion)