bgp path exploration damping (ped) - swincaia.swin.edu.au/talks/caia-talk-110818a.pdfbgp path...

BGP Path ExplorationDamping (PED)

Mattia [email protected]

Centre for Advanced Internet Architectures (CAIA)Swinburne University of Technology

Outline

Introduction

Motivation

Path Exploration Damping - PED

Experimental resultsReduction of update loadMRAI and PED convergence time compared

Conclusions and future work

IETF-81 http://www.caia.swin.edu.au [email protected] 28 July, 2011 2

mailto:[email protected]

http://www.caia.swin.edu.au


Outline

Introduction

Motivation





About this talk

Selected for the Applied Networking Research Prize (ANRP)based on a peer-reviewed paper:

Paper published in the IEEE Journal on Selected Areas inCommunications (JSAC), October 2010 1

A Technique for Reducing BGP Update Announcementsthrough Path Exploration Damping.

Geoff Huston, Mattia Rossi, Grenville Armitage

Project sponsored by the Cisco University ResearchProgram Fund

1http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5586440




http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5586440



What is Path Exploration Damping?

Path Exploration Damping → PEDAlgorithm intended to replace MRAI and RFDMethods to reduce update churn and convergence time inBGPBGP is the de-facto standard for inter-domain routing

BGP – Border Gateway ProtocolMRAI – Minimum Route Advertisment IntervalRFD – Route Flap Damping


PED implementations

CAIAImplemented in QuaggaPatch available for Quagga 0.99.13Download at http://caia.swin.edu.au/urp/bgp/tools.html

Cisco (by Mohammed Mirza)Implemented in CISCO IOS-XE Experimental Version 15.1Running on Cisco ASR1002 (2RU)Currently tested at APNIC Pty. Ltd., Australia




http://www.quagga.net

http://caia.swin.edu.au/urp/bgp/tools.html



Outline

Introduction

Motivation





Simplified BGP speaker design

Router 1

Router 2

Decision process

Router 3

Router 4

Buffer,Filter

Buffer,Filter

Buffer,Filter

Buffer,Filter

Adj-RIB-in RIB Adj-RIB-out

RIB: Routing Information Base – Routing TableAdj-RIB-(in,out): Adjacency RIB (in,out)






Basic BGP dynamics

AS1

AS2 AS3 AS4

AS5

Observer

Prefix: 1.0.0.0/8


Basic BGP dynamics

AS1

AS2 AS3 AS4

AS5

Path: 1

Observer

Prefix: 1.0.0.0/8






Basic BGP dynamics

AS1

AS3 AS4

AS5

AS2

Path: 1

Path: 2,1

Path: 2,1

Observer

Prefix: 1.0.0.0/8


Basic BGP dynamics

AS1

AS4

AS5

AS2

Path: 1

Path: 2,1

Path: *2,1 3,2,1

Path: 3,2,1

AS3

Observer

A: 5,2,1

Prefix: 1.0.0.0/8






Basic BGP dynamics

AS1

AS4

AS5

AS2

Path: 1

Path: 2,1

Path: *2,1 3,2,1 4,3,2,1

Path: 3,2,1

AS3

Observer

A: 5,2,1

Prefix: 1.0.0.0/8


Path Exploration

AS1

AS4

AS5

AS2

Path: 2,1

Path: *2,1 3,2,1 4,3,2,1

Path: 3,2,1

AS3

Path: W

Observer

A: 5,2,1---

Prefix: 1.0.0.0/8






Path Exploration

AS1

AS4

AS5

Path: W

Path: W

Path: *3,2,1 4,3,2,1

Path: 3,2,1

AS3

W

AS2W

Observer

A: 5,2,1---

Prefix: 1.0.0.0/8


Path Exploration

AS1

AS4

AS5

AS2

Path: W

Path: *4,3,2,1

Path: W

AS3

W

W

A

Path: W

Observer

A: 5,2,1---A: 5,3,2,1

Prefix: 1.0.0.0/8






Path Exploration

AS1

AS4

AS5

AS2

Path: W

Path: W

Path: W

AS3

W

A

Path: W

Observer

A: 5,2,1---A: 5,3,2,1A: 5,4,3,2,1

Prefix: 1.0.0.0/8


Path Exploration

AS1

AS4

AS5

AS2

Path: W

Path: W

Path: W

AS3

W

Path: W

Observer

A: 5,2,1---A: 5,3,2,1A: 5,4,3,2,1W

Prefix: 1.0.0.0/8






What is Path Exploration?

An update sequence lengthening the AS-path gradually untilstability is reached


MRAI and RFD

Need to decrease BGP chattiness and Path ExplorationMinimum Route Advertisment Interval – MRAI (RFC 1771)

Apply 30s (default) delay on announcements

Router 1

Router 2

Decision process

Router 3

Router 4

Buffer,Filter

MRAITimer

Buffer,Filter

MRAITimer







Minimum Route Advertisment Interval – MRAI

AS1

AS2 AS3 AS4

AS5

Observer

Prefix: 1.0.0.0/8



AS1

AS2 AS3 AS4

AS5

Path: 1

30 sec.

Observer

Waiting 30s

Prefix: 1.0.0.0/8







AS1

AS3 AS4

AS5

AS2

Path: 1

Path: 2,1

Path: 2,1

30 sec.

Observer

Waiting 60s

Prefix: 1.0.0.0/8



AS1

AS4

AS5

AS2

Path: 1

Path: 2,1

Path: *2,1 3,2,1

Path: 3,2,1

AS3

30 sec.

Observer

Delay 0-90sA: 5,2,1

Prefix: 1.0.0.0/8







AS1

AS4

AS5

AS2

Path: 1

Path: 2,1

Path: *2,1 3,2,1 4,3,2,1

Path: 3,2,1

AS3

30 sec.

Observer

Delay 0-90sA: 5,2,1

Prefix: 1.0.0.0/8



AS1

AS4

AS5

AS2

Path: 2,1

Path: *2,1 3,2,1 4,3,2,1

Path: 3,2,1

AS3

Path: W

Observer

A: 5,2,1---

Prefix: 1.0.0.0/8







AS1

AS4

AS5

Path: W

Path: W

Path: *3,2,1 4,3,2,1

Path: 3,2,1

AS3

W

AS2W

Observer

A: 5,2,1---

Prefix: 1.0.0.0/8



AS1

AS4

AS5

AS2

Path: W

Path: *4,3,2,1

Path: W

AS3

W

W

Path: W

Observer

A: 5,2,1---

Updatedelayed

Prefix: 1.0.0.0/8







AS1

AS4

AS5

AS2

Path: W

Path: W

Path: W

AS3

W

Path: W

Observer

A: 5,2,1------

Prefix: 1.0.0.0/8

Updatedelayed



AS1

AS4

AS5

AS2

Path: W

Path: W

Path: W

AS3

W

Path: W

Observer

A: 5,2,1------W

Prefix: 1.0.0.0/8







AS1

AS4

AS5

AS2

Path: W

Path: *4,3,2,1

Path: W

AS3

W

W

Path: W

Observer

A: 5,2,1---

Updatedelayed

Prefix: 1.0.0.0/8



AS1

AS4

AS5

AS2

Path: W

Path: W

Path: W

AS3

W

A

Path: W

Observer

A: 5,2,1---A: 5,4,3,2,1

Prefix: 1.0.0.0/8







AS1

AS4

AS5

AS2

Path: W

Path: W

Path: W

AS3

W

Path: W

Observer

A: 5,2,1---A: 5,4,3,2,1W

Prefix: 1.0.0.0/8


MRAI and RFD


Apply 30s (default) delay on announcementsMRAI on withdrawals (WRATE) allowed per RFC 4271






MRAI on withdrawals – WRATE

AS1

AS4

AS5

AS2

Path: *1

Path: *2,1 3,1 4,3,1

Path: *3,1

AS3

Path: W

Observer

A: 5,2,1

Prefix: 1.0.0.0/8



AS1

AS4

AS5

Path: W

Path: *1

Path: *2,1 3,1 4,3,1

Path: *3,1

AS3

W

AS2

30 sec.

Observer

A: 5,2,1

Prefix: 1.0.0.0/8







AS1

AS4

AS5

Path: W

Path: *1

Path: *2,1 3,1 4,3,1

Path: *3,1

AS3AS2

Data loss!

Black hole!! Observer

A: 5,2,1

Prefix: 1.0.0.0/8


MRAI and RFD


Apply 30s (default) delay on announcementsMRAI on withdrawals (WRATE) allowed per RFC 4271

Route Flap Damping – RFD (RFC 2439)Flapping = sequence of announcements and withdrawalsSuppress flapping prefixes for 1 hour (or more)






Route Flap Damping – RFD

Router 1

Router 2

Decision process

Router 3

Router 4

RFD Buffer,Filter

RFDBuffer,Filter


Flappingprefix!

W



AS1

AS4

AS5

Path: W

Path: W

Path: *3,2,1

Path: 3,2,1

AS3

W

AS2W

Observer

A: 5,2,1---

Prefix: 1.0.0.0/8







AS1

AS4

AS5

AS2

Path: W

Path: W

Path: 3,2,1

AS3

W

A

Path: W

Observer

A: 5,2,1---A: 5,3,2,1

Prefix: 1.0.0.0/8



AS1

AS4

AS5

AS2

Path: W

Path: *4,3,2,1

Path: W

AS3

W

Path: W

W

Observer

A: 5,2,1---A: 5,3,2,1W

Prefix: 1.0.0.0/8







AS1

AS4

AS5

AS2

Path: W

Path: W

Path: W

AS3

W

A

Path: W

Observer

A: 5,2,1---A: 5,3,2,1WA: 5,4,3,2,1

Prefix: 1.0.0.0/8



AS1

AS4

AS5

AS2

Path: W

Path: W

Path: W

AS3

W

Path: W

Observer

A: 5,2,1---A: 5,3,2,1WA: 5,4,3,2,1W

Prefix: 1.0.0.0/8







AS1

AS4

AS5

AS2

Path: 1

Path: 2,1

Path: *2,1 3,2,1 4,3,2,1

Path: 3,2,1

AS3

RFD suppresses1.0.0.0/8 for 1 hour

Where is my data?

Prefix: 1.0.0.0/8


What are the problems?

MRAI exhibits unpredictable behavior

WRATE creates black holes

RFD penalizes prefix owners even if the misbehaviorhappened further upstream

We want something better!






Outline

Introduction

Motivation





PED algorithm

Delay BGP announcements if the announced AS path islonger than the previously known AS path






PED algorithm illustrated

Router 1

Router 2

Decision process

Router 3

Router 4

Buffer,Filter

PEDITimer

Buffer,Filter

PEDITimer


PEDI: Path Exploration Damping Interval



Router 1

Router 2

Decision process

Router 3

Router 4

Buffer,Filter

per-peer settingper-prefix application

Buffer,Filter

per-peer settingper-prefix application








Router 1

Router 2

Decision process

Router 3

Router 4

Buffer,Filter

announce new best path

no delay

Buffer,Filter

Buffer,Filter


A: 1.0.0.0/8

first announcement /previously withdrawn



Router 1

Router 2

Decision process

Router 3

Router 4

Buffer,Filter

longer path!queue update,

start PEDI timer

Buffer,Filter

Buffer,Filter


update







Router 1

Router 2

Decision process

Router 3

Router 4

Buffer,Filter

shorter path!send update,delete PEDI

timer

Buffer,Filter

Buffer,Filter


update



Router 1

Router 2

Decision process

Router 3

Router 4

Buffer,Filter

withdrawal!send update,delete PEDI

timer

Buffer,Filter

Buffer,Filter


update






Outline

Introduction

Motivation





PED – Data analysis

Experiments using 24 hours of real BGP updatesTwo datasets:

1. APNIC Pty. Ltd., Australia (2 peers)2. University of Oregon – Routeviews, US (5 peers)

AS11537

AS293

AS3727

AS2914

AS48285

Routeviews

AS4777

AS4608

AS6447

AS131702

APNIC






PED – Data analysis

Replayed using the Quagga-Accelerator

BGP cloud Router 1 Router 2

PED Collector

Quagga Accelerator


Outline

Introduction

Motivation









Reduction of update load – APNIC

numberofmessages

0

20k

40k

60k

80k

100k

120k

140k

160k

30sec

MRAI

0sec

MRAI

30sec

PEDI

35sec

PEDI

40sec

PEDI

45sec

PEDI

50sec

PEDI

55sec

PEDI

60sec

PEDI

65sec

PEDI

70sec

PEDI

BGP update m essages

Pref ix announcem entsPref ix w ithdrawals

20%29%


Reduction of update load – Routeviews

numberofmessages

0

50k

100k

150k

30secMRAI

0secMRAI

5secPEDI

10secPEDI

15secPEDI

20secPEDI

25secPEDI

30secPEDI

35secPEDI

40secPEDI

45secPEDI

50secPEDI

55secPEDI

60secPEDI

65secPEDI

70secPEDI

75secPEDI

BGP update m essagesPref ix announcem entsPref ix w ithdrawals

32%18%






Outline

Introduction

Motivation





Convergence time approximation

Control Plane convergence (Optimality)Data plane convergence – Forwarding path (Reachability)

Convergence measured from a single point of view(Optimality)

Router 2

Convergence defined as a route being stable for 5 minutes

240s

Time to converge

30s 30s 30s 120s 5min






Optimality approximation – APNIC data

seco

nds

0

50

100

150

200

400

500

600

700

800

30sM

RAI0s

MRAI

30sPEDI

35sPEDI

40sPEDI

45sPEDI

50sPEDI

55sPEDI

60sPEDI

65sPEDI

70sPEDI

Upper and lower quartileMeanMedian


Optimality approximation – Routeviews data

seconds

0

50

100

150

200

250

400

500

600

700

800

30secMRAI0 sec MRAI5 sec PDI10 sec PDI15 sec PDI20 sec PDI25 sec PDI30 sec PDI35 sec PDI40 sec PDI45 sec PDI50 sec PDI55 sec PDI60 sec PDI65 sec PDI70 sec PDI

Upper and lower quartileMeanMedian






Detailed convergence time analysis

Investigate Reachability vs. OptimalityAnalysis of a whole BGP systemImpact on convergence of 4 events causing instability:

Link failure along the path – alternative path exists

Link recovery

Prefix withdrawal

Prefix announcement


Detailed convergence time analysis

Experimental analysis over 20 testrunsSimple example topology

ASes are single BGP speakersExample: 30s MRAI on all ASes and 35s PEDI on all ASes

Prefix Origin:1.0.0.0/8

AS2 AS3 AS4 AS5 AS6

AS1 AS10 AS11 AS12 AS13

AS20 AS21 AS22






Optimality – Announcement of initial route

Stable System:


* 1 * 2,1 *3,2,1 *4,3,2,16,13,12,11,10,1

*12,11,10,1

*11,10,122,21,20,10,1

*10,1* 1

*13,12,11,10,15,4,3,2,1

*10,1 *20,10,1 *21,20,10,112,11,10,1

AS2 AS3 AS4 AS5 AS6


AS20 AS21 AS22

Data to AS1

Announcement of initial route at AS6:PED: 0 secondsMRAI: 60-120 seconds


Reachability – Link failure

Link failure between AS10 and AS11:


* 1 * 2,1 *3,2,1 *4,3,2,16,13,12,11,10,1

*22,21,20,10,1

*12,22,21,20,10,1* 1

*10,1 *20,10,1 *21,20,10,1

AS2 AS3 AS4 AS5 AS6


AS20 AS21 AS22

Data to AS1

*13,12,11,10,15,4,3,2,1

*12,11,10,1

Reachability achieved (AS11)PED: 0 secondsMRAI: 0-4 or 29-30 seconds






Optimality – Link failure

Link failure between AS10 and AS11:


* 1 * 2,1 *3,2,1 *4,3,2,1

*12,22,21,20,10,1

*22,21,20,10,1

*12,22,21,20,10,1* 1

*5,4,3,2,113,12,22,21,20,10,1

*10,1 *20,10,1 *21,20,10,1

AS2 AS3 AS4 AS5 AS6


AS20 AS21 AS22

Data to AS1

Optimality achieved:PED: 66 seconds (+-jitter)MRAI: 2-58 seconds


Optimality – Link recovery

Link recovery between AS10 and AS11:


* 1 * 2,1 *3,2,1 *4,3,2,16,13,12,11,10,1

*12,11,10,1

*11,10,122,21,20,10,1

*10,1* 1

*13,12,11,10,15,4,3,2,1

*10,1 *20,10,1 *21,20,10,112,11,10,1

AS2 AS3 AS4 AS5 AS6


AS20 AS21 AS22

Data to AS1

Optimality achieved:PED: 0 secondsMRAI: 31-33 and 55-60 seconds






Optimality/Reachability – Prefix withdrawn

Prefix withdrawal at AS1


W W W W

W

WWW

W

W W W

AS2 AS3 AS4 AS5 AS6


AS20 AS21 AS22

Optimality achieved (route withdrawn on every AS):PED: 0 secondsMRAI: 0 seconds


Optimality/Reachability – Prefix (re)-announced

Prefix announcement at AS1


* 1 * 2,1 *3,2,1 *4,3,2,16,13,12,11,10,1

*12,11,10,1

*11,10,122,21,20,10,1

*10,1* 1

*13,12,11,10,15,4,3,2,1

*10,1 *20,10,1 *21,20,10,112,11,10,1

AS2 AS3 AS4 AS5 AS6


AS20 AS21 AS22

Data to AS1

Optimality achieved (same as initial announcement):PED: 0 secondsMRAI: 32-34, 58-60, 76-90 seconds






Outline

Introduction

Motivation





Conclusions

In this talk:PED decreases update loadPED converges to Reachability as fast or faster than MRAIPED converges to Optimality slower than MRAI in one case

In the paper:PED interacts well with MRAIPED can be deployed incrementallyA single PED speaker is beneficial to the BGP system35s PEDI is a safe defaulte value in the MRAI dominatedInternet

In the future:Dynamic PEDI per prefixMore heuristics






bgp path exploration damping (ped) - swincaia.swin.edu.au/talks/caia-talk-110818a.pdfbgp path...

Documents