A Modeling Framework to Understand the Tussle between ISPs ...garetto/slides/perfisp07.pdf · Then...

A Modeling Framework to Understand the Tussle between ISPs and Peer-to-Peer File Sharing Users

IFIP Performance 2007

Michele Garetto – Univ. of TorinoDaniel Figueiredo – Fed. Univ. of Rio de JaneiroRossano Gaeta – Univ. of TorinoMatteo Sereno – Univ. of Torino


An Internet Tale Once upon a time...

user unhappy (“world wide wait”) ISP unhappy (little revenue)

user

InternetISP

Then came broadband access...

user

InternetISP

And they lived happily ever after...

fast!


The Villain Arrives P2P file sharing application

(Kazza, Bittorrent, Emule, etc)

user

users love it! good and free

content, overnight downloads

ISP

ISPs hate it! users using

their link Internet link

utilization gone wild

degrades all subscribers

more bandwidth costs money!

InternetISP


Taking Care of The Villain Mafia

style!

user

seriously threaten application developers!

doesn't seem to work (Napster story)

Is it Really a Villain?

Users love it! Driving force for broadband adoption Increased revenue for ISPs


Some Other Options

user

User unfriendly ideas increase

subscription cost volume based

pricing block / shape P2P

traffic

User friendly ideas acquire more

bandwidth network caching application-layer

redirection

What should the ISP do?

ISP

???


The Real Thing (Data)

P2P represented 60% of InternetTraffic at the end of 2004!


Our contribution Modeling framework to analyze

interactions between P2P file sharing users (their traffic) and ISP economic + performance models

Basic insights about system dynamics

Used to evaluate different strategies to manage P2P traffic


Meet the Players generates

queries quality of service

expectations what's hot,

what's not

user

ISP goal: to make

money! sets subscription

price controls

bandwidth influences P2P

app. behavior P2P application locates object

Network

network architecture

protocols


System Setting

User issues query

Bd constrained resource for ISP

Outside download consumes Bd

number of P2Pusers within ISP

number of P2Pusers outside ISP


Simple System Model

object retrieval probability:

prob. P2P app.locates object

prob.object islocated inside ISP

Model for “Internet to ISP” link

aggregatequery rate

unconstraineddownloads fromwithin the ISP


User Utility Function

Satisfaction model for user i

-1

-0.5

0

0.5

1

0 0.2 0.4 0.6 0.8 1σ

User utility

probability of successful object retrieval

shape parameter

subscription cost

Minimum service level for

user i

user


ISP Utility Function Profit for ISP (revenue - costs)

fixed charge

cost per unit of external bandwidth

revenue from subscribers’ fee

The ISP starts service only if

ISP


Modeling Traffic Locality Probability there exist at least one internal replica

of object replicated r times in the system Number of internal copies

Number of external copies

Probability to download from internal replica

localityparameter


Analytical Results How much bandwidth should the ISP buy

to minimally satisfy the users?

Bmin=max [0,nq min−q r n/N ]

identical users and n >> N

Non-linear behavior (on n) more users, more locality, less BW needed can be zero if n large enough

May not yield profit too few users, too costly to satisfy them

Dependent on multiple parameters


Impact of Object Replication (r)

0

1000

2000

3000

4000

5000

0 10000 20000 30000 40000

Bm

in (

ob

ject

s/d

ay)

Number of users, n

r = 500r = 1000

r = 1500

more replicas, better locality, lower Bmin

more bw neededto support largeruser population

less bw needed(users satisfied locally)


Impact of Subscription Cost (c)

-10000

0

10000

20000

30000

40000

0 5000 10000 15000 20000 25000 30000

c = 0.25c = 1.0c = 1.4c = 1.6

Number of users, n

Uti

lity o

f IS

P

nmin

critical mass of users, nmin

lower cost, more profit earlier, less profit later

ISP doesnot provideservice!


Critical Mass of Users, nmin

0

20000

40000

60000

80000

100000

120000

140000

0 200 400 600 800 1000

nm

in

Average object replication, r

β2 = 6β2 = 5β2 = 4β2 = 3

Cost per unit of bandwidthfor ISP

higher bw cost for ISP, higher critical mass large influence of number of replicas


Model Refinements Simple model

users' access bandwidth are unconstrained

object replication is a parameter

all objects are identical (no popularity)

users availability identical

Refined model relax these assumptions propose object

popularity and replication model


Object Popularity and Replication Model Temporal evolution of object popularity Objects' popularities evolve differently Objects continuously introduced and

removed by users

Analytical technique based on Poisson shot noise process

Number of replicas of an object at time t?


Exampleob

ject

p

opula

rity

time

A video from the news

A popular song

t1 t2

object request

at request time, both have same popularity, but news has more replicas


Limited Bandwidth Refinements

users BW consumptionis limited to b

d

each user within ISP modeledseparately

users upload bandwidthlimited to b

u

rate of downloadrequests to user i


Results from Refined Model

Degenerate to simple model when parameters

set appropriately

Other interesting insights influence of limited upload bandwidth upload/download bandwidth asymmetry object popularity and replication influence of user impatience

0

1000

2000

3000

4000

5000

6000

7000

0 5000 15000 25000

bd = 2000

Bm

in

Number of users, n

bd = 1000bd = 500bd = 100bd = 10


Impact of asymmetric access bandwidths

0

2000

4000

6000

8000

10000

12000

0 500 1000 1500 2000 2500 3000 3500 4000

Upload bandwidth, bu

(for fixed number of users = 20000)

bd = bu

bd = 2 bu

bd = 3 bu

bd = 4 bu

Bm

in (

ob

ject

s/d

ay)

cost for ISP increases as ratio increases better if upload BW is greater than download


Conclusions Development of simple analytical model

economics + performance interaction between P2P users (their traffic)

and ISP insights into strategy for ISP to manage its

traffic Model for object popularity and

replication of independent interest

Future work Multiple ISPs competing with each other


THE END Thank you! Questions? Comments?

CacheLogic Advanced Solutions for P2P Networks

Presentation | WCW 2005

Impact on Service Providers

P2P is thethe dominant protocol In excess of 92% of P2P traffic crosses transit/peering links P2P protocols will aggressively consumeaggressively consume any available bandwidth

capacity Due to P2P’s symmetrical nature on average 80% of upstream

capacity is consumed by P2P P2P affects QoS levels for ALL subscribers Service Providers can not afford to block or restrict P2P ISPs must intelligently manage P2P - blocking and shaping

doesn’t work

P2P is driving consumer broadband uptakeP2P is driving consumer broadband uptake……and broadband is driving P2P uptakeand broadband is driving P2P uptake

HTTP

Other Non P2P

Other P2PFastTrack

eDonkey

BitTorrent

RecognisingGnutella


The ISP perspective vs P2P:threat or opportunity ?

P2P traffic: friend or foe ? friend: driving force for adoption of

broadband access by the users foe: overwhelming amount of traffic

What is the best strategy to manage P2P traffic in my network ? Try to kill it ? Do nothing ? Educate it ? How ?


Strategies to manage P2P traffic Acquire more bandwidth Block P2P traffic Traffic shaping (e.g., priority to non-P2P) Pricing schemes based on user traffic

volumes / bandwidth caps Network caching / customized P2P

application within ISP Application-layer redirection of P2P

traffic


Assumption: n identical users N = 50 millions request rate by user (object/day)

introduction of new contents by

user (object/day)

Results

Minimum required external bandwidth:


The impact of efficacy in exploiting traffic locality ( )

0

20000

40000

60000

80000

100000

0 100000 300000 500000 700000

= 0.25

= 0.5= 0.75

= 1.0

Number of users, n

Minimumrequired bandwidth Bmin

(objects/day)


Model refinements Impact of finite bandwidth of the users

In case of constant traffic load, the cost for the ISP increases if it provides more download bandwidth to the users (!)

The system performance is strongly affected by the upload bandwidth of the users, that should be larger than or equal to the download bandwidth (contrary to common practices, e.g., ADSL lines !)

The impact of bd

0

1000

2000

3000

4000

5000

6000

7000

0 5000 10000 15000 20000 25000 30000

bd = 2000

Bm

in

Number of users, n

bd = 1000bd = 500bd = 100bd = 10

A Modeling Framework to Understand the Tussle between ISPs ...garetto/slides/perfisp07.pdf · Then...

Documents

Transcript of A Modeling Framework to Understand the Tussle between ISPs ...garetto/slides/perfisp07.pdf · Then...