Who Talks to Whom:Who Talks to Whom:Using BGP Data for Scaling Using BGP Data for Scaling

Interdomain Resource ReservationInterdomain Resource Reservation

Ping Pan and Henning Schulzrinne

Columbia UniversityISMA Workshop – Leiden, Oct. 2002

OverviewOverview Reservation scaling

CW: “per-flow reservations don’t scale” true only if every flow were to reserve may be true for sub-optimal

implementations… Based on traffic measurements with

BGP-based prefix and AS mapping looked at all protocols, since too little

UDP to be representative

Reservation scalingReservation scaling Reserve for sink tree, not source-

destination pairs all traffic towards a certain network

destination provider-level reservations

within backbone high-bandwidth and static trunks (but not

necessarily MPLS…) application-level reservations

managed among end hosts small bandwidth and very dynamic flows

Separate intra- and inter-domain reservations

Example protocol design: BGRP

Different growth curvesDifferent growth curves

1

10

100

1,000

10,000

100,000

1,000,000

10,000,000

100,000,000

1,000,000,000

Jan-94 Jan-95 Jan-96 Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02

End Users

Networks

RoutingDomains(AS's)

Estimating the max. number Estimating the max. number of reservationsof reservations

Collected 90-second traffic traces June 1, 1999

MAE West NAP 3 million IP packet headers AS count is low due to short window:

were about 5,000 AS, 60 network prefixes then

May 1999: 4,908 unique source AS’s 5,001 unique destination AS’s and 7,900,362 pairs (out of 25 million)

A traffic snap shot on a A traffic snap shot on a backbone linkbackbone linkGranularity flow discriminators potential

flows

application source address, port 143,243

dest. address, port, proto 208,559

5-tuple 339,245

IP host source address 56,935

destination address 40,538

source/destination pairs 131,009

network source network 13,917

destination network 20,887

source-destination pairs 79,786

AS source AS 2,244

destination AS 2,891

source-destination pair 20,857

How many flows need How many flows need reservation?reservation? Thin flows are unlikely to need resource

reservations Try to compute upper bound on likely

reservation candidates in one backbone router

Eight packet header traces at MAE-West three hours apart on June 1, 1999 90 seconds each, 33 million packets bytes for each

pair of source/destination route prefix destination route prefix

Distribution of connection by Distribution of connection by bandwidthbandwidth

0

10000

20000

30000

40000

50000

60000

70000

< 50bps 50 - 500 (bps) 500 - 2000 (bps) 2000 - 8000 (bps) > 8000 bps

Nu

mb

er o

f C

on

nec

tio

ns

Source-Destination Network Pairs

Destination Networks

The (src-dest / destination) The (src-dest / destination) ratioratio

0

1

2

3

4

5

6

7

< 50 (bps) 50 - 500 (bps) 500 - 2000 (bps) 2000 - 8000 (bps) > 8000 (bps)

Gai

n

ResultsResults Most packets belong to small flows:

63.5% for source-destination pairs 46.2% for destination-only

only 3.5% (3,261) of the source-destination pairs and 10.9% (1,296) of destinations have average bit rate over 2000 b/s thus, easily handled by per-flow reservation

more above-8000 b/s destination-only flows than source-destination flows large web servers?

Aside: Estimating the number Aside: Estimating the number of flowsof flows

In 2000, 4,998 bio. minutes ~ 500 bio

calls/year local (80%), intrastate/interstate toll

15,848 calls/second not correct assumes equal distribution

AT&T 1999: 328 mio calls/day 3,800/second

The Hierarchical Reservation ModelThe Hierarchical Reservation Model

AS-5

(Private Network)

AS-4

NAP

LAN

AS-3

(Transit Backbone)Private Peering

Private Peering

Application-layer reservationApplication-layer reservation

Provider-level reservationProvider-level reservation

AS-2

AS-1

R1R1

R2R2

R3R3

R4R4

LL

ConclusionConclusion Communications relationships

granularity and “completeness” flow distribution

Questions: traffic seems to have changed qualitatively

more consumer broadband, P2P see “Understanding Internet Traffic Streams”

protocol behavior funnel-behavior may differ for QoS candidates e.g., large PSTN gateways but no funnel for (e.g.) media servers