24
P2P-ISP Cooperation: Risks and Mitigation in Multiple-ISP Networks Mung Chiang, Princeton University Aliye Özge Kaya, Wade Trappe WINLAB, Rutgers University {ozgekaya, trappe}@winlab.rutgers.edu

P2P-ISP Cooperation: Risks and Mitigation in Multiple … · Risks and Mitigation in Multiple -ISP Networks ... We model the hints given by the oracle to the P2P ... P2P-ISP Cooperation:

Embed Size (px)

Citation preview

P2P-ISP Cooperation: Risks and Mitigation in Multiple-ISP Networks

Mung Chiang,Princeton University

Aliye Özge Kaya, Wade TrappeWINLAB,

Rutgers University{ozgekaya, trappe}@winlab.rutgers.edu

WINLAB

Overview

Recent P2P-ISP cooperation proposals and their risks Possible outcomes of P2P-ISP cooperation Framework to model the interaction between ISPs and P2P

providers Risk Mitigation

WINLAB

Current P2P-ISP Interaction:No Cooperation: Lose-lose situation for all!

P2P Providers: -Limited capability to infer network topology-Network oblivious peering-Randomize the traffic to prevent filtering P2P traffic by ISP

ISPs: -Increased cost and backbone traffic because of the P2P traffic-Incentives to filter and limit P2P traffic

End-Users

•Poor P2P experience( lower rates, higher delay)

•Incentives to change the ISP and P2P provider

WINLAB

P2P-ISP Cooperation: Win-Win for all?

Peering decisions avoiding bottlenecks, costly connections.

Reduced costs for ISPsImproved performance for P2Ps

P2P takes into account

ISP’s preferences

P2P Provider discloses

peer locations

ISP discloses network related

Information

End-Users

Better P2P performance:

Download from the closer peers

Higher download rates, reduced delay

Satisfied customer

WINLAB

Does P2P-ISP Cooperation bring only benefits in multi-ISP case?

The benefits have been demonstrated in the single ISP case but potential risks have been not well examined for the multiple ISP case

We first look to the existing cooperation proposals, and identify possible risks.

ISP P2PISP P2P

ISP

ISPISP

ISP

P2P

P2P

WINLAB

Recent Proposals for P2P-ISP Cooperation:Approach A: ISP supported oracle [Aggarwal08]

P2P provider suggests the P2P users possible peers

• Preferential Treatment: P2P can manipulate the list such that to promote the peers in some ISP’s more than the others

ISP ranks the peers for the P2P user

• ISP has the full freedom to manipulate P2P-traffic

• P2P provider can use this ISP oracle to infer sensitive Information

P2P users chooses the peers accordingly

• Choices made by ISPs are not always in the interest of the end user

• Worse Performance?

WINLAB

Recent Proposals for P2P-ISP Cooperation:Approach B: CDN supported Oracle [Choffnes08]

Monitor CDN redirection properties

• Peers redirected to the same CDN server are assumed to close to each other

Assign the peers with the similar redirection properties

• Localized traffic might not yield always the best performance for ISPs or P2P providers.

Peer Assignment without any direct involvement of ISP

• What if the ISP reveals the CDN servers in its network or P2P infers it?

WINLAB

Recent Proposals for P2P-ISP Cooperation: Approach C: P4P [Xie08]

ISP reveals the network topology, P2P reveals peer locations

• Exchange inacccuratetopology, misleading information ( e.g. unexistingcongestion).

iTracker assigns each peer to a cluster called PID

• Computes the metric p4p distance between the PIDs

• Peering suggestion to theAppTracker in form of of p4p distances between the clusters

AppTracker assigns peers close in terms of p4p distance to each other

• P2P provider or ISP can reverse engineer the topology by clustering PID’s with similar preferences.

• Ignore hints from the iTracker

WINLAB

Comparison of Approaches A, B,CA: ISP Oracle B : CDN Oracle C: P4P

Manipulation of Traffic by ISP alone

ISP has fullfreedom tomanipulatetraffic.

not possible partially (mightmisrepresent itsown network)

Can the P2Pprovider infernetworktopology?

partially (useISP oracleas trafficmonitor)

partially (CDNsreplica servesas identifiers ofISPs)

partially (clusterPID’s with similarpreferences,iTracker as trafficmonitors)

Preferential Treatment by P2P (hiding-promoting peers)

possible(manipulatethe list ofthe suggestedpeers)

possible (useCDN Replicaservers asidentifiers forthe preferredISPs)

possible(misrepresentlocation of thepeers, ignorehints fromiTracker)

Manipulation of Traffic by ISP alone

Can the P2Pprovider infernetworktopology?

Preferential Treatment by P2P (hiding-promoting peers)

WINLAB

Outcomes of P2P-ISP Cooperation

Benefit Drawback Counter-Intuitive

P2P Efficiency(Higher rates, lessdelay)

Lose privacy and sensitive information like topology

Unfairness (betterbenefits for non-cooperatingISP)

Reduced inter-ISPcost

Difficult to verify theexchanged information

Selfish behaviorleading degradationof the benefits

Reduced bandwidthconsumption

Benefit Drawback Counter-intuitive

WINLAB

Drawbacks of Cooperation

Use “No Valley and Prefer Customer Policy” of routing also as giving preference for the destination– Customer ISPs will be forced to

send much more content to the provider ISPs

Exchanging Misleading Information– Inaccurate topology, nonexisting

congestion etc. Preferential Treatment

– ISPs or P2Ps selectively release information in a way to favor some P2P providers or ISPs over others.

WINLAB

The cooperating ISPs may not be willing to send content to the peers in non-cooperating ISPs

Counter-Intuitive Outcomes:Better benefits for the non-cooperating ISPs

Average Rate [R]

ISP A ISP B ISP C

No Cooperation 0.6 0.6 0.6

All Cooperating 0.75 0.75 0.75

ISP C not coperating (Enough Peers) 0.75 0.75 1

ISP C not cooperating (Not enough peers)

0.75 0.75 0.5

• Cooperating ISP : Downloads content with probability ½ and rate R within the network with probability ¼ from any other network at rate ½ R • Non cooperating ISP: Downloads content with probability 1/3 from any network

No Cooperation

All Cooperating

ISP C not cooperating (Enough Peers)ISP C not cooperating (Not enough peers)

WINLAB

What do the P2P-ISP cooperation frameworks have in common? Can we investigate the risks in unified framework?

The common entity in all approaches is an oracle which gives hints to the P2P provider. The main difference is whether ISP gets individual and indirect involvement (Approach A) No involvement (Approach B) Direct Involvement (Approach C)

in oracle P2P does the peering assignments based on the hints from the

oracles ISP downloads content more from some ISPs than the others as

a result of these peering assignments.

WINLAB

Overview of our framework:

Consider ISP j request a traffic volume V, each ISP i supplies of it.

We model the hints given by the oracle to the P2P provider by a metric called .

We model the manipulations by the P2P provider by a metric . P2P provider determines which ISP provides what fraction of V

to the ISP j based on and

βα

α β

Vijθ

WINLAB

Modeling hints from oracles

: Willingness of ISP i to send content to ISP j : Willingness of ISP i to receiver content from ISP j : Joint preference of ISP i to send content to ISP j

(Compromise by oracle)

ISP i ISP j5.0=ijβ

1=sijβ 0=r

ijβsijβrijβ

ijβ

WINLAB

P2P has the freedom to ignore or enforce preferences of some ISPs (captured by metric ) more than others.

The metric assigned by the P2P provider to manipulate the traffic flow from ISP i to ISP j.

Modeling manipulations of P2P providers

ijβ

ijα

WINLAB

Traffic Allocation by the P2P provider for ISP j

Let be the fraction of the traffic demand V of ISP j supplied by ISP i

P2P provider does the traffic allocation by solving:

Solution mathematically equivalent to the water filling

ijθ

WINLAB

Manipulations by the P2P provider

ISP 9 is less preferred P2P provider manipulates traffic

from ISP 7 and ISP 8 to ISP 9 We set

equal to We vary between 0 and 2.

98899779 ,,, αααα

)(1 jiij ≠=β2=iiβ

αα

WINLAB

Net Income versus α• As P2P provider observes the preference of ISP 7 and ISP 8 to send much more content to ISP 9, ISP 9 is enforced to download from ISP 8 and ISP 9.

The P2P provider puts more weight to the preferences of ISP 7 and ISP 8

As the P2P observes the preferences of ISP 7 and ISP 8 more the cost for

ISP 9 increases

WINLAB

When ISP joins the cooperation the ISP lets the P2P provider know the minimum fairness levels to continue the cooperation.

Rate Fairness and Cost Fairness: P2P can raise the fairness level for that ISP by reallocating

some of the traffic at a higher rate or lower cost. If the P2P provider reallocates using the strategy

that results in average rate then fairness level will be raised to . .The P2P provider determines from:

Risk mitigation by enforcing fairness

WINLAB

Rate fairness

As the P2P observes the preferences of ISP 7 and ISP 8 to send and receive more content from ISP 9 the rate

fairness decreases for ISP-9

WINLAB

Guideline Principles for P2P-ISP Cooperation

Avoid ISP owned oracles for inter-ISP peering decisions Enforce fairness between ISPs and P2P providers by auditing

by third parties or punishment strategies Observe the hints from the oracles partially, make some

portions of the peering assignments based randomly or some other inferred metric like delay.

WINLAB

Concluding Remarks

We focus this work on the risks of cooperation, especially when multiple ISPs co-exist and form their own peering relationships.

We built a mathematical model for the three recent approaches of cooperation, classifying the potential risks and metrics, and construct numerical examples and analytical results on some of the major unintended or counter-intuitive behaviors.

We believe that P2P-ISP cooperation is generally helpful to all parties involved, and can have significantly good practical impact after their risks are discovered and bounded.

Quantifying of the guideline principles is an interesting next step in fully understanding the best P2P-ISP collaboration mechanism

WINLAB

THANKS!

[email protected][24]