1

How location diversity ate traffic engineering’s cake

Abhigyan, Aditya Mishra, Vikas Kumar,Arun VenkataramaniUniversity of Massachusetts Amherst

2

Examples:◦ CDNs◦ P2P applications◦ Mirrored

websites◦ Cloud computing

Location diversity

Location diversity: Ability to download content from multiple locations

3

ISPs have several objectives, e.g., minimizing congestion, decisions about upgrading link capacity

ISPs optimize link utilization based metrics. e.g. maximum link utilization (MLU)

Traffic engineering

4

Traffic engineering and location diversity

Traffic engineering (ISPs)

Location diversity (CDNs)

Internet

traffic

5

How do TE schemes compare accounting for location diversity in the Internet?

Problem

1. Introduction2. Motivation

1. Location diversity and traffic engineering2. Metric of comparison

3. Evaluation4. Conclusion

Outline

Application adaptation to

location diversity

Traffic matrix

New Routing

Location diversity changes TE problem

Traffic engineering

Content demand

8

Location diversity changes TE problem

100

Mbp

s,

0.1m

s

100 Mbps,

10ms

1

23

100

Mbp

s,

10m

s

10 Mb x 10 req/s = 100 Mbps

10 Mb x 5 req/s = 50 Mbps

OSP

F W

t = 2

OSP

F W

t =

150 M

bps + 5

0

Mbp

s50 M

bps

Maximum link utilization ( MLU )= 1

OSPF W

t = 1

9

Location diversity changes TE problem

100

Mbp

s,

0.1m

s

100 Mbps,

10ms

1

23

100

Mbp

s,

10m

s

OSP

F W

t = 2

OSP

F W

t =

150 M

bps + 5

0

Mbp

s50 M

bps

OSP

F W

t = 1

25 M

bps + 2

5

Mbp

s25 M

bps + 2

5

Mbp

s

Expected MLU = 0.5

25 Mbps +

25Mbp

s

MLU = 0.75

10 Mb x 10 req/s = 100 Mbps

10 Mb x 5 req/s = 50 Mbps

OSPF W

t = 1

Location diversity changes TE problem (2)

Location diversity increases capacity

100

Mbp

s

100 Mbps

1

23

10 Mb x 10req/s = 100 Mbps

100 Mbps 10

0 M

bps

10 Mb x 20req/s = 200 Mbps

Increase in capacity = 200/ 100 = 2

11

1. Motivation1. Location diversity and traffic engineering2. Metric of comparison

2. Evaluation3. Conclusion

Outline

Without location diversity◦ Capacity = 1/MLU

12

MLU poor metric of capacity

100

Mbps

1

23

25 Mbps

25 M

bps

100

Mbps

MLU = 0.25

100

Mbps

Capacity = 100/25 = 4

100 Mbps

max supportable demand current demand

Capacity =

Without location diversity◦ Capacity = 1/MLU

With location diversity◦ Capacity >= 1/MLU

13

MLU poor metric of capacity

100

Mbps

1

23

30 Mbps

100

Mbps

25 M

bps5

Mbps

MLU = 0.25

180 Mbps

90 M

bps9

0 M

bps

Capacity > 180/30 = 6

Need a new metric to quantify capacity under location diversity

max supportable demand current demand

Capacity =

Surge protection factor (SPF) SPF = Maximum

supportable surge (linearly scaled) in traffic demand

14

SPF = 200/30 = 6.66

100

Mbps

1

23

30 Mbps

100

Mbps

25 M

bps5

Mbps 1

00

Mbps

100

Mbps

200Mbps

15

SummaryLocation diversity significantly impacts TE

1. Capacity increases2. Capacity (SPF) not captured by 1/MLU

16

1. Introduction2. Motivation3. Evaluation

1. TE schemes2. Measuring SPF3. Capacity results (SPF)

4. Conclusion

Outline

TE schemes comparedTE Schemes

(Almost online) optimal TE [OPT]

(Offline) “optimal” TE using MPLS [MPLS]

(Offline) TE using OSPF link weight optimization [OptWt]

(Offline) Multi-TM optimization TE [COPE]

(Oblivious) Static shortest path routing with inverse-capacity link weights [InvCap]

18

Measuring SPF

Is demand satisfied

?

Increase demand by Δ

SPF = demand/(initial demand)

Demand = initial demand

YES

NO

19

Capacity results (SPF)

MPL

S

Opt

Wt

COPE

InvC

ap0

0.25

0.5

0.75

1

No Loc DivLoc Div = 2Loc Div = 4

SP

F c

om

pare

d t

o

OP

T

InvCap worst case No LocDiv = 50% sub-OPT

LocDiv = 30% sub-OPT

1. All TE schemes achieve near-optimal capacity with location diversity.

2. Even no TE scheme is at most 30% sub-optimal with location diversity.

20

“How location diversity ate traffic engineering’s cake”◦ Any TE scheme performs the same as Optimal TE.◦ No TE scheme performs at most 30% worse.

Conclusions