PAM 2008 1

A Measurement Study of Internet Delay Asymmetry

Abhinav Pathak Purdue UniversityHimabindu Pucha Purdue UniversityYing Zhang University of MichiganY. Charlie Hu Purdue UniversityZ. Morley Mao University of Michigan

PAM 2008 2

RTT: FWD + REV

FWD

REV

RTT is easy to measureFWD + REV = RTT

PAM 2008 3

Round Trip Time (RTT) is a widely used network metric in server/peer selection CDN

Select closest replica Overlay multicast

Choose a suitable parent/child in the tree Internet distance prediction

Calculate proximity from a landmark

De’ Facto Internet Delay Metric: RTT

PAM 2008 4

Often Times, One-Way Delay is More Relevant Examples scenarios

Multicast streaming applications Real-time interactive applications

Multi-player games Internet distance prediction Understand routing performance

PAM 2008 5

OWD measurement requires Access to both ends

No daemon in OS Strict time synchronization

Result: OWD is approximated as half of RTT

But OWD Measurement is Hard

Conventional Wisdom• Delay is symmetric• FWD = REV

PAM 2008 6

Outline: Questions to Answer

Does delay asymmetry exist?

What are reasons for delay asymmetry?

How dynamic is delay asymmetry?

PAM 2008 7

Tools and Testbed

Tools Owping: Implements one way active

measurement protocol (RFC 4656) Paris-Traceroute

Testbed Planetlab 180 GREN, 25 Commercial nodes Trace collection: 10 days in April 2007

PAM 2008 8

Owping Relies on Time Synchronization NTP – Two parameters

Clock drift Relative error to a NTP server Applied to timestamp on each host

Error estimate Added up to report overall error estimate

PAM 2008 9

Server Client

owampd(resource broker)

owpingControl

Connection setup

owampd(control)

fork

owampdTest Endpoint

fork

Request Results

owpingTest Endpoint

fork

OWD Test packets

How Does Owping Work

Source: http://e2epi.internet2.edu/owamp/

PAM 2008 10

NTP Drift and Error Estimate

40% of nodes have Error Estimate > 20 ms

PAM 2008 11

Trace Pruning

Remove trace if

NTP error estimate > 10ms Leaves 82 GREN nodes and 12 commercial nodes

Sum of NTP error estimate > 3% of RTT for a node pair Leaves primarily long distant routes

PAM 2008 12

GREN Vs Commercial nodes

Planetlab mostly contains GREN nodes GREN – Global Research and Education Network

GREN to GREN (G2G) path properties are different from Commercial to Commercial (C2C)

G2C and C2G properties are close to C2C

- On the Impact of Research Network Based Testbeds on Wide-area Experiments [Pucha et. al. – IMC06]

We consider G2C-C2G-C2C paths only

PAM 2008 13

Delay Asymmetry – FWD/RTT

Delay fraction for all node pairs

PAM 2008 14

Delay Asymmetry – Absolute Values

Y = (1/2) X(conventional

wisdom)

RTT = 150 msFWD = 100 msREV = 50 ms

PAM 2008 15

Questions to answer

Does delay asymmetry exist? Yes

What are reasons for delay asymmetry?

How dynamic is delay asymmetry?

PAM 2008 16

Reasons for Asymmetry in OWD What are reasons for delay asymmetry

Temporary congestion in forward or reverse path Transient events

Forward and reverse paths are different Path Asymmetry

Can we correlate delay and path asymmetry? Measure path using traceroute Need a metric to quantify path asymmetry

PAM 2008 17

Metric for Path Asymmetry – Path Similarity Coefficient AS level path asymmetry

A = {All ASes in forward path} B = {All ASes in reverse path}

AS path similarity coefficient = |A B| / |A B|Ո Ս

Router level path asymmetry A = {All routers in forward path} B = {All routers in reverse path} Router path similarity coefficient = |A B| / |A B|Ո Ս

PAM 2008 18

Path Asymmetry

70% paths haveAS level

Similarity coeff. > 0.6

20% paths haveRouter level

Similarity coeff. > 0.6

PAM 2008 19

Delay Asymmetry Vs Router Level Path Similarity Coefficient

Router Level Path Similarity Coefficient

Delay fraction ~ 0.5 when router-level

asymmetry nears unity

Delay fraction fluctuatesbetween 0.3 to 0.7

PAM 2008 20

Questions to answer

Does delay asymmetry exist? Yes

What are reasons for delay asymmetry? Observed good correlation with path asymmetry

How dynamic is delay asymmetry?

PAM 2008 21

Dynamics of Delay Asymmetry

PAM 2008 22

Correlating FWD change and RTT change

D

A

C B

PAM 2008 23

Correlating Routing Events and Delay Asymmetry Change Measurement setup

Traceroute and owping all nodes

Repeat after every 20 minutes

Calculate path change Inter AS / Intra AS

Measure reverse path at the same time

PAM 2008 24

Correlation Results

80% of Intra AS path change cause FWD to

change by < 10 ms

80% of Inter AS path change cause FWD to

change by < 20 ms

PAM 2008 25

Delay Dynamics – Observations Intra AS path change

More frequent to observe Most of the times path changes in both directions Fwd & Rev delays change simultaneously

Inter AS path change Less frequent Two cases

Only fwd AS path changes Both fwd and rev AS paths change

Delay change is larger in magnitude

PAM 2008 26

Questions to answer

Does there exists delay asymmetry? Yes

What are reasons for delay asymmetry? Observed good correlation with path asymmetry

How dynamic is delay asymmetry? Depends on inter/intra AS path change

PAM 2008 27

Summary Methodology

Measuring OWD Pruning strategy

Based on error estimates provided by NTP

Measurement results Considerable levels of delay asymmetry Delay asymmetry is dynamic

RTT could change due to FWD change or REV change or both

Analyzing the cause Weak correlation between router level asymmetry and delay

asymmetry Delay asymmetry dynamics

Inter/Intra AS route change effects delay asymmetry differently

PAM 2008 28

Thank You

Questions?