DYNAMICS OF PREFIX USAGE AT AN EDGE ROUTER Kaustubh Gadkari, Dan Massey and Christos Papadopoulos 1.

DYNAMICS OF PREFIX USAGE AT AN EDGE ROUTERKaustubh Gadkari, Dan Massey and Christos Papadopoulos

1

2

Outline

Introduction – BGP RIB and FIB growth Motivation Methodology Results Conclusions and future work

3

Outline

Introduction – BGP RIB and FIB growth Motivation Methodology Results Conclusions and future work

4

Inter-Domain Routing

CSU

129.82/16, 129.82/18 …

Google

74.125/16, …

BGP

5

BGP Tables

Router Memory

Routing Table(RIB)

Routing Updates Routing Updates

Forwarding Table(FIB)

Line Card Line Card Line Card

(Prefix, Path)

Forwarding Table(FIB)

Forwarding Table(FIB)

(Prefix, Outgoing Interface, Next

Hop)

6

Outline

Introduction – BGP RIB and FIB growth Motivation Methodology Results Conclusions and future work

7

Table Growth

The size of the global routing is growing rapidly.

Increasing routing table size requires more memory on routers.

Operators are forced into faster upgrade cycles.

http://bgp.potaroo.net/as6447/

8

Implications of Table Growth

RIB size affects FIB size.

FIB scaling is arguably more important. FIBs stored on line card memory, which is

smaller and more expensive than main memory.

9

Outline

Introduction – BGP RIB and FIB growth Motivation Methodology Results Conclusions and future work

10

Motivation

Some current research focuses on scaling the FIB by storing partial forwarding information.

But how to choose prefixes in a reduced FIBs? Prefixes receiving most packets should probably

be in the table. But are there better criteria to define a

dominant set? How does it behave?

11

Motivation (contd.)

Previous work focused on prefix popularity over days or weeks. Tradeoff that prefers ease of selection.

Optimum prefix selection is a hard problem. Factors: traffic volume, activity patterns and

interplay of traffic dynamics

This work: understand prefix dynamics better.

12

Outline

Introduction – BGP RIB and FIB growth Motivation Methodology Results Conclusions and future work

13

Methodology

Monitor links to two tier-1 provider links (1Gb/s each) at a regional ISP.

Dataset: two simultaneous 24-hour packet traces from these links.

Use outgoing packets only.

14

Old Metric: Global Rank

Rank prefixes according to number of packets during the full 24-hour trace. This is the prefix’s

global rank.

Plot number of prefixes that account for a given fraction of packets. Results

corroborate previous studies.

15

Measuring Prefix Dynamics

Split traffic trace into small intervals (5 min).

Rank prefixes in each interval according to number of packets. We call this the prefix’s interval rank.

16

New Metrics

Duty cycle Fraction of the total number of intervals in

which the prefix receives at least one packet.

Mean rank difference Variation of a prefix’s interval rank from its

global rank.

17

Outline

Introduction – BGP RIB and FIB growth Motivation Methodology Results Conclusions and future work

18

Duty Cycle

Measure of prefix’s activity.High duty

cycle, high

traffic

Low duty

cycle, low

traffic

~200 prefixes

~56,000

prefixes

High duty

cycle, low

traffic

~24,000

prefixes

Low duty cycle, high

traffic

~10 prefixes

Global Prefix Rank

19

Duty Cycle Observations

Popular prefixes have high duty cycles. Always get packets.

Several popular prefixes have a duty cycle of > 90%.

20

Mean Rank Difference

Measure of prefix’s “busy-ness”.

Stable Prefixes

Less than 1% of all prefixes,

40% of all packets

Generally Popular Prefixes5% of all prefixes, 55% of

all packets

Generally Unpopular Prefixes

60% of all

prefixes, 5% of all

traffic

Global Prefix Rank

21

Outline

Introduction – BGP RIB and FIB growth Motivation Methodology Results Conclusions and future work

22

Conclusions

Understanding dynamic behavior of FIB prefixes is important for reduced-FIB designs.

Proposed two new metrics: Duty cycle and mean rank difference We corroborated previous work showing set of

dominant prefixes is small. New metrics characterize the dominant set better,

which is generally active and busy. Majority of the prefixes have very low activity

opening up caching opportunities. Results encouraging in terms of developing

reduced FIB designs.

23

Future Work

Investigate prefix dynamics at other ISPs.

Investigate prefix dynamics at different intervals.

Develop a design for efficient forwarding using a reduced FIB.