Computer Networks 56 (2012) 1118–1130

Contents lists available at SciVerse ScienceDirect

Computer Networks

journal homepage: www.elsevier .com/locate /comnet

Cooperative traffic management for video streaming overlays

Konstantin Pussep a,⇑, Frank Lehrieder b, Christian Gross a, Simon Oechsner b,Markus Guenther a, Sebastian Meyer a

a Multimedia Communications Lab, Technische Universität Darmstadt, Germanyb Institute of Computer Science, University of Würzburg, Germany

a r t i c l e i n f o

Article history:Available online 17 December 2011

Keywords:Peer-to-peerISPsLocalityTraffic managementVideo streaming

1389-1286/$ - see front matter � 2011 Elsevier B.Vdoi:10.1016/j.comnet.2011.10.027

⇑ Corresponding author. Address: TU Darmstadtdeturmstraße 10, D-64283 Darmstadt, Germany. Tefax: +49 6151 16 6152.

E-mail address: pussep@kom.tu-darmstadt.de (K

a b s t r a c t

Peer-to-peer (P2P) based overlays often ignore the boundaries of network domains andmake traffic management challenging for network operators. Locality-aware techniquesare a promising approach to alleviate this impact, but often benefit only network operatorsand fail to provide similar benefits to end-users and overlay providers. This is especiallysevere with video streaming overlays that are responsible for a large amount of Internettraffic. In this paper we present and evaluate a collaborative approach where a networkoperator measures the behavior of overlay users and promotes a subset of them in termsof up- and download bandwidth. This creates an incentive to users and overlay providersto cooperate by using locality awareness according to the operator’s policies. We evaluateour approach both with a real application and via extensive simulations to analyze the userselection metrics and the impact on different network operators. Our study shows that thiscooperative traffic management approach leads to a situation that is beneficial for users,content providers, and network operators.

� 2011 Elsevier B.V. All rights reserved.

1. Introduction

Video streaming is emerging as one of the most popularservices in the Internet. It comprises a large set of differentapplications ranging from video communications over livestreaming to video-on-demand (VoD). According to theCisco Visual Networking Index [1], video traffic contributestoday already more than 33% of the total consumer Inter-net traffic and this fraction is predicted to grow to 91%by 2014.

Video applications constitute a major challenge for con-tent providers and network operators since they require ahigh amount of network resources in order to deliver a goodquality-of-experience and to satisfy the users’ expectations.

. All rights reserved.

, FB 18, KOM, Run-l.: +49 160 7658865;

. Pussep).

A promising solution to this problem is peer-assisted videostreaming where users watching a video also upload partsof the video to other users in a peer-to-peer (P2P) basedmanner. This reduces the load on the video servers andmakes those systems more scalable in terms of the numberof concurrent users that can be served [2–4].

Due to the high amount of traffic produced by videoapplications, the need for a reasonable management ofsuch traffic arises for network operators. The most promi-nent solution for that problem is locality awareness, whichis currently under discussion in the IETF working group onapplication layer traffic optimization [5]. Locality aware-ness equips peers with the knowledge about the topologyof the underlying physical network and/or preferences ofthe operator to enable a ‘‘better-than-random’’ peer selec-tion for data exchange. This approach has been shown tobe highly beneficial for the network operators [6,7] sincea large fraction of the costly inter-domain traffic can besaved. However, there is no clear benefit of applying local-ity awareness for the users and the content providers. Inmost scenarios, the performance from the users point of

K. Pussep et al. / Computer Networks 56 (2012) 1118–1130 1119

view is not improved [6,8] or may decrease depending onthe concrete implementation of locality awareness andoverlay characteristics [9]. This is a severe drawback,which might also slow down or even inhibit an Internet-wide deployment of such mechanisms.

This problem is addressed by a measurement-basedoptimization approach that we proposed in Pussep et al.[10]. Its intention is that operators yield some of their ben-efits to network-friendly users that apply locality aware-ness. Those users, which we call highly active peers(HAP), can be identified by the operator through networkmeasurements. To reward them for being network-friendly, the operator can offer faster Internet access forthem, i.e., increase their upload and download capacities.Content providers also profit in this case since their userscan share more traffic among themselves, which relievesthe load on the servers. Therefore, content providers havean incentive to implement locality-aware mechanisms intheir clients. In summary, this traffic management scheme,which is based on voluntary cooperation of all involvedparties, facilitates a triple-win situation for network oper-ators, users, and content providers.

In this paper, which is an extended version of Pussepet al. [10], we refine our original proposal, present a proto-type implementation, and conduct a comprehensiveperformance evaluation of our cooperative traffic manage-ment approach based on simulations and testbedexperiments with a prototype. Our evaluation uses amesh-based VoD overlay and comprises the HAP promo-tion as well as locality awareness mechanisms. It showsthat the combination of HAP promotion and localityawareness indeed leads to the desired triple-win situationfor network operators, content providers, and users. Forthat purpose, we extend our previous work [10] in thefollowing ways. (1) We propose and compare differentmetrics, which can be measured by network operators, toselect peers to be promoted to HAPs. Their performanceis evaluated in a simulation study comprising 50,000 peersand 2,500 video files. (2) We introduce a prototypicalimplementation of the proposed HAP promotion mecha-nism in the German Lab testbed and assess the impact ofthe traffic management on all three stakeholders. To thatend, we adapt locality awareness mechanisms for file-sharing to video streaming and implement them in theTribler streaming client [11] that we used for our experi-ments. (3) As shown in our previous work the HAPpromotion mechanism might lead to an undesired increasein outbound traffic, especially for early adopters. We intro-duce a technique to avoid such an increase and demon-strate its effectiveness. (4) Finally, we compare theperformance of HAP promotion to other managementapproaches for P2P traffic such as blocking and caching.

The paper is structured as follows. Section 2 presentsbackground information and discusses related work. InSection 3, we introduce our cooperative traffic manage-ment approach. The first part of our performance evalua-tion based on testbed experiments with real clients in asingle swarm is presented in Section 4. In Section 5, we ex-tend the scenario to a large number of swarms and inves-tigate it via simulations. Finally, we conclude our paper inSection 6.

2. Background and related work

In this section we describe the state-of-the-art for thetraffic management of overlay applications. Then, we re-view P2P-based VoD services.

2.1. Overlay traffic management

The largest incentive to apply overlay traffic manage-ment exists for network operators. Their goal is to reduceinter-domain traffic, which may be costly due to inter-provider agreements. Therefore, the earliest traffic manage-ment approaches, namely traffic shaping or blocking, areunilateral mechanisms implemented by the network opera-tors. However, these measures reduce the performance ofthe overlay and lead to dissatisfied users [12]. An alternativeapproach, which does not reduce the overlay performance, isthe use of P2P caches. These caches can serve the requests oflocal peers and reduce incoming traffic in this way [13].However, this advantage comes at the price of the resourcesthe network operator has to provide for the caches, the effortto support different protocols, and may also raise legalissues.

Finally, a traffic management approach that has re-ceived much attention recently is locality awareness[14,15]. Here, the overlay structure is modified to prefershort, local connections over longer and more costly in-ter-domain connections. The required information can beobtained via an operator provided oracle [7] or an iTracker[16], be derived from the DNS redirection behavior of con-tent distribution networks as proposed in [17], or be basedon BGP routing tables [18]. At the P2P client this can beimplemented by the so-called Biased Neighbor Selection(BNS) [6], as well as by Biased Unchoking (BU) [8]. Theformer changes the mechanism for establishing overlayconnections, the latter the selection of peers to uploadcontent to. BNS and BU have been developed primarilywith BitTorrent in mind and therefore work with anyBitTorrent-based overlay, such as the overlay we use inour performance evaluation. Results from prior evaluationsof these mechanisms show that they can reduce inter-domain traffic significantly under certain conditions, whilethe performance is not reduced from an average user’spoint of view [19]. However, under realistic conditions asubset of the users might experience negative effects fromlocality awareness [9,20]. Thus, there are no clear incen-tives for an end-user to use a locality-aware system.

In this work, we compare the effect of locality aware-ness with a new approach, namely the assignment of addi-tional capacity to highly active peers, and show that thecombination of these approaches is beneficial for all threestakeholders.

2.2. P2P video streaming overlays

The approach of a P2P overlay supporting dedicatedvideo-on-demand streaming servers and its potential for loadreduction on these servers have been evaluated in the recentpast [21]. However, the focus is commonly on the benefits for

1120 K. Pussep et al. / Computer Networks 56 (2012) 1118–1130

the content provider, without taking into account the effectson traffic management and on the end-user.

A BitTorrent Assisted Streaming System for VoD (BASS)is one of the early approaches that extends BitTorrent tosupport streaming [2]. In the investigated scenario the nec-essary streaming server bandwidth can be reduced by onethird through the support of a P2P overlay. In Vlavianoset al. [3] present and evaluate an enhanced BitTorrent vi-deo streaming protocol named BiToS. Toast is another Bit-Torrent-based support scheme for VoD servers aspresented in Choe et al. [4]. Similar to BASS, it shifts loadfrom the dedicated VoD servers to a BitTorrent overlay,with parts of the video that cannot be delivered by theoverlay being requested from the server. Again, the focusis on overlay mechanisms and the impact on the networkis neglected. The main performance indicator used forthe evaluation is the load reduction for the VoD server(up to 90%), i.e., its upload traffic savings by utilizing theoverlay network [4]. In Abboud [22] another approach forpeer-assisted VoD is presented. The system includes ad-vanced piece selection and upload allocation techniquesespecially suitable for a commercial scenario.

Tribler [23] is an open-source P2P VoD streamingplatform which is based on BitTorrent, similar to [2,3].In contrast to these, a usable client application exists,which facilitates validation of various extensions in a realprototype. Tribler uses a BitTorrent variant, called give-to-get, that replaces both the chunk selection and the peerselection algorithms to cope with the challenges of videostreaming [24]. For the chunk selection, the chunk set afterthe current playout position is organized into three prioritysets, where the high priority set is filled in-order and thesubsequent sets according to the rarest-first policy. Thework in [25] uses this protocol to demonstrate how toallocate a set of servers efficiently, so that a streamingperformance can be guaranteed to users while minimizingthe server load.

Applying locality awareness to video streaming workssimilar to the BitTorrent overlay. The common idea is toprefer local peers over remote peers. This distinction canbe applied, for example, when a peer decides which otherpeers to connect to or from which peers to request certaindata. Several works presented locality-aware techniquesapplicable to (live or on-demand) streaming overlays[26–29]. Most of them consider the case of live streamingbased on application-layer multicast trees and propose tobuild subtrees according to the network boundaries, end-to-end delays, or similar metrics [27,28]. Another approachis presented by Wang et al. where the upload bandwidth isallocated considering (among other parameters) the net-work topology [26]. While being able to reduce the inter-domain traffic these algorithms cannot improve thestreaming quality or reduce server load, thus, missing anincentive for the overlay to apply them.

3. Measurement-based traffic management of VoDoverlays

The previous sections revealed that none of thepresented approaches has achieved a solution that is

acceptable for all three parties: network operators, overlayproviders, and users. For example, any strategy for trafficmanagement that explicitly throttles inter-domain trafficresults in degradation of the overlay performance. On theother hand, client-side locality awareness brings no benefitto the overlay and, therefore, lacks an incentive to bewidely adopted by overlay providers and users. The reasonis that the performance of P2P-based content deliveryoverlays relies on two factors: availability of the contentand the upload bandwidth of participating peers. Whilethe overlay performance only considers these factors glob-ally, the network operators should attempt to improvethem in the local domain. Only then can they successfullypromote locality without hurting overlay performance.

In order to overcome the aforementioned limitations,we propose that network operators provide a clear andmeasurable incentive for locality-aware behavior at theoverlay level. To this end, the network operators shouldpromote the selected users compliant with overlay andnetwork requirements by increasing their upload anddownload bandwidth. This approach is able to increasethe overlay performance and to reduce the amount of in-ter-domain traffic resulting in the so-called triple-winsituation:

1. Network operators can reduce the costly inter-domaintraffic generated by P2P overlays. At the same time theydo not risk disgruntling their clients.

2. Overlay providers benefit from the increased band-width of overlay peers. Their own resources are relievedfrom traffic load even more than without any activetraffic management. Thus, the overlay provider canchoose either to save server hosting costs, to increaseQoS for the delivered content, or even to increase thevideo resolution.

3. Users can directly and indirectly benefit from incentive-based traffic management by the network operator. Theprovision of additional resources directly improves theperformance of the overlay and, therefore, indirectlythe experienced quality of each overlay user. Further-more, the users selected for promotion can benefit fromfree resources without extra fees by using network-friendly applications (or application configurations).

A simplified example is shown in Fig. 1. Inside of its do-main the network operator monitors the usage profiles ofpeers to identify those having the highest (potential) im-pact on the overlay traffic (step 1). The network operatormanages this data in its billing and provisioning systemthat aggregates the historical data and offers it to the net-work management system (step 2). The network manage-ment system analyzes this data, selects best peers andassigns them additional bandwidth, which makes themmore attractive for other peers (step 3). Other local peersshould detect these promoted peers and download contentfrom them (step 4).

Next, we discuss in detail how single componentsshould be implemented: Monitoring of user traffic, selec-tion of peers to promote, and the actual promotion of peersto HAPs. We also present a mechanism to avoid the (poten-tially) undesired increase in outbound inter-domain traffic.

3. Promote

2. Historical data

4. Preferred connection

Heavyuser

Casualuser

User traffic

NetworkManagement

System

BillingSystem

Rest of theInternet

Inter-domain links

Casual user

Operator's control traffic

Local network domain

Peeringrouter

1. Collect statistics

Fig. 1. Example of highly active peer promotion from the perspective of a single network operator.

K. Pussep et al. / Computer Networks 56 (2012) 1118–1130 1121

3.1. Monitoring usage statistics

In order to make the right decisions, the network oper-ator should monitor the parameters listed in Table 1 in reg-ular intervals that have the same duration s as thepromotion interval. An important property of the moni-tored parameters is the lack of content-awareness, whichis desired by the network operator (because of the possiblecopyright infringements) and the overlay provider (be-cause of the possible DRM issues).

This data is gathered in regular intervals and can be col-lected either by the means of the in-network monitoring(such as NetFlow [30]) or from the overlay, as discussedin [10]. We propose to collect data from the networkequipment in the first place. The non-forgeability of thisdata is its major advantage. Furthermore, this does not re-quire any changes to the monitored overlay applicationsand, therefore, supports any P2P overlay.

Due to the protocol-independent nature of our ap-proach the network operators are not required to classifythe traffic according to a specific P2P protocol (such as Bit-Torrent, PPLive etc.) but can rely on statistics that containthe total traffic of the given user. Since network operatorsare able to perform volume-based charging for their users,they have to know their upload and download traffic,which serve as a good indicator to distinguish peer-typefrom client-type users [31]. Additionally, network opera-tors that already make use of deep packet inspection(DPI) or traffic pattern detection [32] can classify theexchanged data according to the protocol type: P2P,HTTP, etc. (but not necessarily the exact protocol in use).

Table 1Traffic statistics collected per user and measurement interval t.

Parameter Description

Vup(t) Traffic uploaded

Vlocup ðtÞ Traffic uploaded within the local domain

u(t) Upload bandwidth during last measurement interval tVdown(t) Traffic downloaded

Vlocup ðtÞ Traffic downloaded within the local domain

Therefore, in the remainder of this study we assume thatthe network operator is able to identify P2P users andconsider only those P2P users in our evaluations.

Furthermore, network operators have a global view onthe locality of Internet addresses that belong to their do-main, which allows them to distinguish between localand remote traffic. A real-life example from a networkoperator monitoring facility is shown in Fig. 2 and visual-izes the total upload and download traffic of a single peer.For instance, the active link usage during the morninghours can be attributed to a P2P (file sharing) application.

3.2. Selection of highly active peers

The goal of the selection algorithm is to identify mostactive and available peers based on the measured data(see Table 1). These peers must be able to act as locality-promoting HAPs and bias the overlay traffic for morelocality.

In order to select suitable peers, we consider the follow-ing types of relevant user behavior: overlay contribution,seeding ratio, and network-friendliness as discussed be-low. We define HAP selection metrics that assign a ratingvalue R(t) to each peer p for the currently examined periodt so that the network operator can select the highestranked peers to be promoted.

The Contribution metric C(t) represents total contribu-tion of the peer to the overlay. The contribution of thepeers is crucial for the overlay performance, since peerswho already contributed a lot, could potentially contributeeven more. The metric is calculated as:

CðtÞ ¼ VupðtÞuðtÞ � s

with u(t) – 0 and s > 0 being the duration of the measure-ment interval. The normalization by u(t) � s makes the met-ric agnostic to the customer’s bandwidth and permits a faircomparison between HAPs and normal peers.

The Network-friendliness metric F(t) prefers peers thatprobably apply locality-aware peer selection or that are

Fig. 2. Example of a traffic monitoring of a single user over 24 hours.

1122 K. Pussep et al. / Computer Networks 56 (2012) 1118–1130

popular inside the domain, being more quickly discoveredby other local peers. We compute F(t) as:

FðtÞ ¼Vloc

up ðtÞþVlocdownðtÞ

VupðtÞþVdownðtÞif VupðtÞ þ VdownðtÞ > 0;

0 otherwise:

(

The Seeding ratio metric S(t) is the amount of a peer’supload traffic relative to the total traffic volume. It enablesselecting peers that are altruistic in the sense that they stayonline to provide content and not only to consume it (thisallows us to avoid the measurement of actual online timeof peers, that is practically rather difficult). We calculateS(t) as1:

SðtÞ ¼SðtÞ ¼ VupðtÞ

VupðtÞþVdownðtÞif VupðtÞ þ VdownðtÞ > 0;

0 otherwise:

(

Furthermore, we include a random HAP selection metricthat assigns a random rating value to a peer. It serves as abaseline for comparison since it does not rely on anyknowledge about the peers’ behavior.

As discussed above, ideally the HAP promotion shouldconsider both network- and overlay-related metrics. Thequestion arises how to combine such metrics. One possiblesolution is a weighted sum approach as proposed, e.g., in[10,27]. However, the weighting of such different metricsas presented above is rather difficult. Instead, we proposea hierarchical filtering approach, where one metric is usedto remove unsuitable peers and the remaining candidatesare ranked by the second metric. Thereby, we use Contri-bution C(t) as our primary metric and Network-friendlinessF(t) as our secondary metric.2

Let us consider a set P of candidate peers N and the pri-mary metric C. We define Cr1 ðPÞ as the r1th percentile of Paccording to C. In the first filtering step, the best r1 � N peersare selected as: PC ¼ fp 2 PjCðpÞP Cr1 ðPÞg. This allowsremoval all peers that do not show significant activity inthe overlay. In the second filtering step, the best r2 � jPCjpeers are selected resulting in the overall number of of

1 Note that our definition differs from the seeding ratio common in theBitTorrent networks which is defined as VupðtÞ

VdownðtÞ. Our version has the benefitof normalization, since all values will be in the range [0:1]. Besides this, theresulting order of peers is the same.

2 in Section 5.2 we demonstrate that Contribution outperforms Seedingratio as a single HAP selection metric.

r1 � r2 � N selected peers. Using Network-friendliness as thesecondary metric (and its r2th percentile) we obtainPC;F ¼ fp 2 PC jFðpÞP Fr2 ðPCÞg. For example, if 20% of peerscan be promoted, the operator could set (r1,r2) = (0.5,0.4).

The HAP selection algorithm must also consider that thebehavior of peers and, therefore, the measured values,might fluctuate over time, either due to random and diur-nal effects, or because of changes in the behavior of singleusers. The usage of historical data can alleviate the effect ofsuch fluctuations, but older values should receive lowerweights to allow for long-lasting changes in user behaviorand in order to keep the entrance barrier low for newpeers.

Therefore, we choose the modified exponential movingaverage [33] to aggregate historical data with exponen-tially diminishing weights. Given a certain selection metricR for the current measurement interval t 2 N, this resultsin the aggregated rating R0 with:

R0ðtÞ ¼ a � RðtÞ þ ð1� aÞ � R0ðt � 1Þ for t > 0;0 for t ¼ 0:

�

A useful property of this metric is that the networkoperator has to store only a single historical value per peer,resulting in a reduced storage and computational complex-ity. The value of a depends on the interval duration s andthe user behavior fluctuations. For our evaluation, we seta = 0.5, so that an old value R(t � x) is effectively weightedwith 1/2x.

3.3. Changing user access profile

The proposed incentive-based traffic management tech-nique requires on-the-fly automated updates of the cus-tomer access profiles. In our case this is the increase ofthe totally accessible upload and download bandwidth ofcertain users. While there are different operator- and ven-dor-specific solutions possible, we present one real exam-ple where our approach was successfully implemented bya network operator.3

The resulting network architecture is presented inFig. 3. In this case study, the customer access bandwidthis not limited by a DSLAM but it is throttled by the use of

3 http://www.prime-tel.com.

Internet

DSLcustomers

Peeringrouter

DSLrouter T c

shaper Billing &Provisioning

Fig. 3. Case study of HAP promotion.

K. Pussep et al. / Computer Networks 56 (2012) 1118–1130 1123

a customized Linux-based traffic shaper. Thereby, thebandwidth provided to single users by the DSLAM is oneor two times higher than the shaped bandwidth. Such anarchitecture that is otherwise used to provide triple-playor similar services that require QoS enforcement for differ-ent types of traffic, can be easily reused for our approach.Here, the billing and provisioning subsystem stores thecurrent bandwidth assigned to single peers and the trafficshaper enforces these limits. The applied limits are thenreconfigured dynamically to promote certain customersto HAPs. This case study demonstrates the feasibility ofthe proposed approach, where the reconfiguration of cli-ents takes place at least once per day (though it could bealso done in the range of few hours).

4 http://www.german-lab.de/.

3.4. Avoiding outbound traffic increase

So far, we considered increasing the users’ bandwidthglobally, i.e., the network operator cannot control how cli-ents use the additional bandwidth. Indeed, peers promotedto HAPs might upload a lot of traffic to remote peers andeven increase the outbound inter-domain traffic. Even ifsuch peers might be demoted after the next HAP selectionround, the new set of peers might exhibit the same behav-ior. The reasons for such network-unfriendly behaviormight be diverse, including the lack of locality awarenessin the overlay or a deficit of available bandwidth (or de-sired content) in other network domains. As a conse-quence, promotion of peers might lead to an undesiredincrease of the outbound traffic.

To avoid this, we propose that the network operatorincreases the bandwidth only locally, that is, inside of itsnetwork domain. This is an alternative to the global band-width increase that we consider as the default option. Thisdecision may have a significant impact on the overallperformance of the HAP promotion approach. A localbandwidth increase is more efficient from the networkoperator’s perspective if outbound traffic is costly for thegiven operator. In the scenario presented in Section 3.3,

the local bandwidth increase can be enforced by the trafficshaper by matching the receivers IP address with the localIP prefixes.

4. Evaluation of the single swarm case

The goal of this study is to assess whether the involvedstakeholders can benefit from the HAP promotion in a real-istic environment. For this purpose we use an instru-mented client on the basis of the Tribler VoD client [11].We perform the measurements in the German Lab testbed4

and apply a real-time user promotion. This section coversthe results of the following experiments: (1) Impact of adop-tion by all network operators, which compares the impact oflocality awareness and HAP promotion on users, overlayproviders, and network operators; (2) Impact on an earlyadopter, which is the case when only a single network oper-ator deploys HAP promotion.

4.1. Methodology

The scenario for our experiments is shown in Table 2.Due to the restrictions of the testbed environments, welimit the experiments both in terms of the overlay sizeand the number of considered parameter variations. Theserestrictions are dropped in the subsequent simulationstudy (see Section 5). Our setup comprises four networkdomains: three customer domains containing peers andone server domain. The overlay peers are equally distrib-uted among the network domains and, if locality aware-ness is in place, only 50% of peers make use of it. Everytestbed configuration is repeated three times and we pres-ent average values and standard deviations.

We capture the following metrics to understand the im-pact of the mechanisms under study: traffic uploaded byservers to capture the costs for the overlay provider (andimplicitly for the overlay users) and inter-domain trafficto capture network operators’ costs. For the latter metricwe also distinguish between the inbound and outboundinter-domain traffic since they might have different impacton costs for certain operators. Content providers are repre-sented using a prototype implementation of adaptive serv-ers [25]. Those servers reside in a separate domain anddata exchange between peers and servers counts asinter-domain traffic. We further capture the total playbackdelays of individual users.

4.1.1. Locality-aware peer selectionWe integrated the locality-aware peer selection mecha-

nisms into the Tribler client with the primary focus on flex-ibility. Our instrumented client uses a client-side variationof Biased Neighbor Selection (BNS) [6] and Biased Uncho-king (BU) as described in [8]. Upon connection establish-ment, e.g., when the peer discovers new potentialneighbors via the tracker response, peer selection accord-ing to BNS takes place. Thus, a peer connects preferentiallyto peers located within the same network domain. Uponupload allocation, when a peer selects a neighbor for the

Table 2Setup of single swarm experiments.

Parameter Values (default, variations)

Experiment duration 90 min.Videos Length = 15 min, bitrate = 540 kbpsTopology 3 customer domains and 1 server domainAvailable servers 0–10 (adaptive)Server bandwidth (upload) 2048 kbpsPeer bandwidth (down, up) 2048/kbps, 512 kbpsPeer arrival rate Exponentially distributed with 8 peers/min.Peers’ online time 75% selfish and 25% altruistic- Selfish tonline ¼ tvideo � z; z ¼ Uð0;1Þ ðuniformÞ- Altruistic tonline ¼ tvideo þ z; z ¼ exp 1

video length=2

� �Locality awareness none, 50% of peers- BNS filtering 90% local connectionsAdoption of HAP promotion none, all, single operator- Peer promotion interval 5 min.- Ratio of promoted peers 20% per domain- Additional bandwidth per HAP 100%

1124 K. Pussep et al. / Computer Networks 56 (2012) 1118–1130

optimistic unchoke, peer selection according to BU takesplace. Here, a peer tries to unchoke a random local peer.

Both strategies use an internal peer selection API thatranks candidate peers based on various sources. Thesesources can be internal knowledge databases such as pre-loaded GeoIP databases5 or external knowledge databasessuch as an ALTO compliant service [5]. Peer ratings fromexternal knowledge databases are cached for a configurableamount of time in order to reduce communication overheadand to avoid additional delays during connection establish-ment and unchoking. This architecture allows easy adoptionof new knowledge databases. It further facilitates the usageof the implementation in various environments, such astestbed experiments or real-world deployments.

4.1.2. HAP promotionNetwork operators deploy HAP promotion in order to

identify beneficial peers according to the combined HAPselection metric (see Section 3.2) and modify their up-stream capacities. Due to the restrictions of the underlyingtestbed, we cannot implement HAP promotion at the net-work level by e.g. using the tc suite.6 Instead, we emulatethe rate limitation on the client-side. First, the upstreambandwidth of each client is set to a predefined access profileaccording to testbed parameters. During runtime operatorscollect bandwidth utilization reports from instrumented cli-ents and instruct selected clients to update their internalrate limits according to their current promotion status.

4.2. Impact of HAP promotion by all operators

In order to measure the impact of our approach, fourdifferent combinations of traffic management approachesare considered: (1) no traffic management (No TM), (2)locality-aware peer selection without HAP promotion(Loc), (3) all network operators apply HAP promotionbut peers do not behave locality-aware (HAP), and (4)

5 http://www.maxmind.com/app/ip-location.6 http://www.linux.org/docs/ldp/howto/Traffic-Control-HOWTO/

intro.html.

combination of HAP promotion and locality awareness(HAP + Loc).

Fig. 4 presents the outcome of HAP promotion appliedby all operators for the involved stakeholders (users, over-lay provider, and network operators). We observe(Fig. 4(a)) that locality-aware peer selection achieves alarge reduction of inter-domain traffic (40% of inboundand 55% of outbound traffic). This reduction goes even fur-ther if we combine locality awareness with HAP promo-tion. As expected, the No TM case shows the largestamount of server traffic and locality-aware peer selectiondoes not reduce this server traffic significantly (Fig. 4(a)).On the other hand, HAP promotion (with or without local-ity awareness) results in server traffic reduction by 36–38%.

Additionally, Fig. 4(b) shows the performance of local-ity-aware and oblivious users in two different scenarios.We observe that locality awareness decreases streamingperformance in terms of total delay. In contrast, HAP pro-motion increases the streaming performance and can com-pensate for negative effects that are introduced by usinglocality-aware peer selection. The graph also shows thatthere is almost no difference between the performance ofoblivious users in both scenarios. We further observe thatmostly locality-aware peers benefit from local HAPs.

Furthermore, our measurement data reveals that withinpromoted peers there were four times more locality-awarethan oblivious peers, which means that locality-awarepeers are rewarded in terms of additional free bandwidth.This provides an incentive to be network-friendly even ifthe performance might be slightly degraded compared tothe oblivious peers. To conclude this experiment, we canobserve that in cases when all network operators applyHAP promotion the combination with locality awarenessachieves the best situation for the overlay provider andthe network operators.

4.3. Impact of adoption by a single operator

In this experiment we consider the case when only asingle operator applies HAP promotion. In this regard, only

Fig. 4. Single swarm scenario: impact of HAP promotion by all operators.

K. Pussep et al. / Computer Networks 56 (2012) 1118–1130 1125

peers inside of its network domain can be promoted toHAPs and only peers inside of its domain apply locality-aware peer selection. Out of those peers, 20% can be pro-moted to HAPs and 50% behave locality-aware.

The outcome of this scenario is shown in Fig. 5. Evenwith only 50% of peers behaving locality-aware, the earlyadopter clearly benefits from the combination of HAP pro-motion and locality awareness. As Fig. 5(a) shows, both theinbound and outbound traffic of operator 1 are roughly40% lower than for the other operators (despite the sameamount of peers per network domain). Furthermore,regarding the server traffic we observe that with HAP pro-motion applied only by operator 1 the server traffic de-creases (cf. Fig. 5(b)), though not at the same extent aswhen all operators deploy HAP promotion.

We also compared the streaming performance for usersbelonging to the early adopter and other operators but ob-served no significant difference. Regarding the users, weobserved negligible deviations between different operatorsregarding the streaming experience. At the same time,mostly locality-aware peers benefit from HAP promotionby receiving additional bandwidth (more than 80% of pro-moted peers are locality-aware).

Fig. 5. Single swarm scenario: impact of HAP prom

5. Evaluation of the multi swarm case

The previous section demonstrated the effect of HAPpromotion with a real application in a testbed environ-ment. However, due to the limitations of the testbed, wedid not study the effect of longer promotion intervals (forat least a few hours) that can be realized in today’s net-works (see Section 3.3). Therefore, in this section we applysimulations as a means to assess the impact of such pro-motion intervals with large amount of users participatingin multiple swarms. We analyze the impact of differentparameters of our approach and compare HAP promotionwith other traffic management approaches.

5.1. Methodology

Our simulation scenario models the behavior of a com-mercial peer-assisted VoD overlay that spans the domainsof several network operators. An efficient implementationof this system in a custom discrete event-based simulatorallows us to simulate a scenario of four weeks.

Table 3 shows the simulation setup (with default valuesshown in bold). The peers are spread over ten heterogeneous

otion by a single operator (early adopter).

1126 K. Pussep et al. / Computer Networks 56 (2012) 1118–1130

network domains following the distribution of Germanusers collected by Ookla [34]. We further model a separatedomain for the overlay provider, which contains only thecontent servers. The domains are organized in a star topol-ogy, which is sufficient to capture the inter- and intra-do-main traffic of single operators.

We model the user behavior based on various measure-ment studies revealing heterogeneous usage patterns [35–37,31]. For this purpose, we divide the users into fourgroups based on two independent properties: seedingbehavior and inter-request time. For the inter-requesttime, we apply the Pareto distribution to user requests,resulting in 20% heavy users generating 80% of the stream-ing requests (similar to the patterns observed by Cho [31]and Basher [38]). Since the seeding time describes the levelof altruism, we divide the users into selfish users, that don’tseed at all, and altruistic users, whose seeding times areexponentially distributed with a mean of one hour. Thecontent is modeled with 2,500 videos where each videohas an average bitrate of 950 kbps. We model the contentpopularity by letting users choose videos to watch accord-ing to a Zipf distribution (with Zipf parameter a = 0.85)[39].

The simulation model also includes all relevant mecha-nisms of HAP promotion (monitoring, selection, and band-width update). By default, the increased bandwidth is notrestricted to the local domain (i.e., global increase). The pro-motion interval is set to one day and 25% of peers are pro-moted by default. Independent of the HAP promotion,peers can behave either network-oblivious or locality-awarewhen choosing their communication partners. In mixedscenarios, only a certain percentage of peers behave local-ity-aware. Since servers are considered as most costly fromthe overlay perspective, both peer selection policies preferremote peers over content servers when choosing down-load sources.

While HAP promotion and locality awareness at theoverlay level are the main traffic management (TM)techniques analyzed in this paper, we also consider twoalternative mechanisms available to network operators asa baseline for comparison: blocking of inter-domain P2P

Table 3Setup for experiments with multiple swarms.

Parameter Values (d

Peers 50,000- Initial upload bandwidth 700 kbp- Cache size 4 GB- Seeding behavior 50%: 0 h- Inter-request time 20%: 8 h- Peer selection policy locality-Videos 2,500- Size 800 MB- Rate 950 kbp- Popularity distribution Zipf (a =Network Operators 10 custo- Size distribution 32, 17, 1HAP promotion 25% of p- HAP selection metric contribu- Total additional bandwidth 100%, 0–- Bandwidth increase global, lAlternative mechanisms no TM, b

traffic and deployment of P2P caches that act as superpeersbut serve only local peers.

We capture the same metrics as in the single swarmexperiments (see Section 4.1) with the exception of theplayback delays since in the simulation model the serversassure that the required video data is supplied in time.We further focus on the total inter-domain traffic that isthe sum of inbound and outbound traffic.

5.2. General impact of HAP promotion

This scenario contains heavy users and casual userswho can be either altruistic or selfish. However, we assumethat all peers behave locality-aware. This assumption isdropped in the subsequent experiment.

Fig. 6 demonstrates the impact of the total bandwidthassigned for HAP promotion and its interplay with HAPselection metrics in use. Even with the random HAP selec-tion metric the inter-domain traffic can be significantly re-duced by up to 40% (see Fig. 6(a)). At the same time theserver traffic can be reduced from 280 to 45 TB (reductionof 85%), which results in the intended win–win situation(see Fig. 6(b)). Besides the random selection metric theother metrics perform even better, resulting in higher ben-efits both for the overlay and network. The contributionmetric turns out to be the best HAP selection metric inall setups, reducing the inter-domain traffic by up to 55%and server traffic by up to 90%. If we consider the operatorsinvesting only 100% of additional upload bandwidth, theimprovement is still significant, with 43% inter-domaintraffic reduction and 88% server traffic savings with thebest metric. This increase of the total upload bandwidthby 100% is considered as our working point for the nextexperiments.

In order to understand the reason that the contributionmetric outperformed the other metrics, Fig. 8 presents thepercentage of promoted peers within single behavioralgroups (for a single simulation run). We observe that theHAP selection based on peers’ contribution prefers heavyusers over casual, but also prefers altruistic behavior inthe second place. These users are frequently online and

efault,variations)

s

and 50%: 1 hand 80%: 48 haware, oblivious, mixed

s0.85)mer and 1 server domain3, 12, 12, 4, 3, 3, 1, 1%eers each 24 htion, network-friendliness, seeding ratio, random300%

ocallocking, P2P caches

Fig. 7. Combined vs. single HAP selection metrics.

K. Pussep et al. / Computer Networks 56 (2012) 1118–1130 1127

are able to serve more requests from other users, and, evenmore, they request more content and, therefore, can servemost popular content. For comparison, the seeding ratiometric prefers altruistic users that stay longer online toserve other users, but their request frequency is often low-er than that of heavy users.

5.3. Incentives for locality awareness

In the previous experiment we saw that if 100% of peersbehave locality-aware, then the contribution metric pro-vides the highest benefit both for network operators andthe overlay provider. In this experiment, we drop theassumption of a fully locality-aware overlay and vary thepercentage of locality-aware peers between 0% and 100%.

We study the effect of the network-friendliness metricand its combination with the contribution metric (seeFig. 7). When the percentage of locality-aware peers isequal to zero, all metrics perform the same, while theincreasing locality awareness degree amplifies the effectof HAP promotion. Furthermore, the contribution metricagain outperforms other metrics and the random selectionperforms worst. Finally, the combination of the contribu-tion and network-friendliness metrics offers a trade-off be-tween pure contribution and pure network-friendliness,both for the overlay and the network (cf. Fig. 7(a)). We fur-ther consider the impact of HAP selection metrics on theserver traffic and observe that the combined metric pro-vides a reduction comparable to the contribution metric,thus outperforming the random selection and pure net-work-friendliness (see Fig. 7(b)).

In order to understand the impact on users, we plot thepercentage of promoted locality-aware peers as a functionof their total percentage in the overlay. (Fig. 9). We observethat with both selection metrics based on the network-friendliness, the probability to be promoted is higher forlocality-aware users because their percentage in the pro-moted peers is higher than their percentage in the overlaypopulation. Thus, users have a clear benefit to behave

Fig. 6. Impact of single HA

locality-aware, which in turn results in a higher benefitfor the network operators.

P selection metrics.

Fig. 8. Type of users promoted by single HAP selection metrics.

Fig. 9. Locality-aware peers within promoted peers.

1128 K. Pussep et al. / Computer Networks 56 (2012) 1118–1130

5.4. Alternative solutions and the effect of local bandwidthincrease

In this experiment we compare HAP promotion withalternative mechanisms to traffic management (TM).Fig. 10 shows the performance of blocking, P2P caches,

Fig. 10. Comparison with alternative

HAP promotion (with global or local bandwidth increase),and no traffic management at all. While each of this mech-anisms can be applied by network operators, we also ana-lyze the impact of different locality awareness degrees atthe overlay level. Without HAP promotion (the no TMcurve), we observe only a minor benefit of locality-awarepeers for operators and no impact on the server traffic. Ifnetwork operators apply blocking of inter-domain P2Pconnections, the inter-domain traffic can be reduced by35%. However, this savings for the network operators comeat the cost of server traffic increasing by a factor of two.Therefore, this approach does not achieve the desiredwin–win situation. On the other hand, P2P caches andHAP promotion show the desired behavior regarding theimpact of higher locality awareness at the overlay level.The lowest inter-domain traffic is achieved with a high de-gree of locality-awareness. Furthermore, HAP promotionwith local bandwidth increase (as proposed in Section 3.4)reduces the server traffic by 38% providing a clear incen-tive for the overlay to behave locality-aware, which in turnbenefits the network operators (cf. Fig. 10(a)). We concludethat both P2P caches and HAP promotion result in the de-sired win–win situation, while the latter approach avoidsthe legal issues and is protocol-independent. To show thepotential of P2P caches we plot both caches with infiniteresources (cacheinf) and with realistic upload bandwidthof one gigabit per second and one terabyte storage space(cache1Gbps1TB). In addition, a P2P cache serves all localpeers and consequently does not provide an incentive forthe users to act according to the operator’s policies.

6. Summary and conclusions

This paper presented a cooperative approach to trafficmanagement of P2P-based streaming overlays, calledhighly active peer (HAP) promotion. Instead of unilateralapproaches such as traffic shaping or overlay-side localityawareness, all stakeholders contribute to obtain mutualbenefit. The overlay peers behave locality-aware byexchanging data preferentially with local peers, while the

traffic management solutions.

K. Pussep et al. / Computer Networks 56 (2012) 1118–1130 1129

network operators promote selected users in terms of theiraccess bandwidth. This creates an incentive for the overlay(both the overlay provider and users) to employ locality-aware mechanisms. In addition, the network operatorsare able to decrease the inter-domain traffic.

To show the feasibility, we presented a proof-of-con-cept study performed by a real network operator with realcustomers. This comprises in particular the measurementof the traffic exchanged by individual users and the abilityto modify users’ access profiles dynamically. We proposeddifferent metrics for network operators to select peers tobe promoted to HAPs and investigated their performance.

To this end, we evaluated our approach in a testbed andthrough large scale simulations. Our analysis showed thatinter-domain traffic and the load on servers of the contentproviders can be reduced significantly. Furthermore, local-ity-aware end-users receive the benefit of additional ‘‘free’’bandwidth and, contrary to pure locality awareness, ob-serve almost no penalty in terms of streaming perfor-mance. We also considered the HAP promotion by only asingle operator and demonstrated that even in that casethis operator benefits. This shows that HAP promotioncan be applied successfully even if only some networkoperators support it. Finally, we compared HAP promotionwith different approaches to overlay traffic managementand observed our approach performing similarly to thebest alternative (P2P caches with high capacity). However,it lacks the associated content-awareness as well as proto-col-dependence and provides incentives for users andoverlay providers to behave locality-aware at the sametime.

Acknowledgements

This work has been funded partially in the frameworkof the EU ICT Project SmoothIT (FP7-2007-ICT-216259)and by the Federal Ministry of Educations and Researchof the Federal Republic of Germany (Förderkennzeichen01 BK 0800, GLab). The authors alone are responsible forthe content of the paper.

References

[1] Cisco visual networking index: Forecast and methodology, 2009–2014 (2010). URL <http://www.cisco.com/en/US/solutions/collateral/ns341/ns525/ns537/ns705/ns827/white_paper_c11-481360.pdf>.

[2] C. Dana, D. Li, D. Harrison, C.N. Chuah, Bass: Bittorrent assistedstreaming system for video-on-demand, in: IEEE Workshop onMultimedia Signal Processing, 2005.

[3] A. Vlavianos, M. Iliofotou, M. Faloutsos, Bitos: Enhancing bittorrentfor supporting streaming applications, in: IEEE INFOCOM, 2006, pp.1–6.

[4] Y.R. Choe, D.L. Schuff, J.M. Dyaberi, V.S. Pai, Improving vod serverefficiency with bittorrent, in: ACM MULTIMEDIA ’07, 2007, pp. 117–126.

[5] Charter of the IETF Working Group on Application-Layer TrafficOptimization (ALTO), 2011, <http://www.ietf.org/html.charters/alto-charter.html>.

[6] R. Bindal, P. Cao, W. Chan, Improving Traffic Locality in BitTorrent viaBiased Neighbor Selection, in: International Conference onDistributed Computing Systems (ICDCS), 2006.

[7] V. Aggarwal, A. Feldmann, C. Scheideler, Can ISPS and P2P userscooperate for improved performance?, ACM SIGCOMM ComputerCommunication Review 37 (2007) 29–40

[8] S. Oechsner, F. Lehrieder, T. Hoßfeld, F. Metzger, K. Pussep, D. Staehle,Pushing the performance of biased neighbor selection through

biased unchoking, in: IEEE International Conference on Peer-to-Peer Computing (IEEE P2P), 2009.

[9] F. Lehrieder, S. Oechsner, T. Hoßfeld, D. Staehle, Z. Despotovic, W.Kellerer, M. Michel, Mitigating Unfairness in Locality-Aware Peer-to-Peer Networks, International Journal of Network Management(IJNM), Special Issue on Economic Traffic Management 21 (1)(2011) 3–20.

[10] K. Pussep, S. Kuleshov, C. Groß, S. Soursos, An incentive-basedapproach to traffic management for peer-to-peer overlays, in:Workshop on Economic Traffic Management (ETM), 2010.

[11] Tribler Client. [link]. URL <http://tribler.org/>.[12] P. Eckersley, F. von Lohmann, S. Schoen, Packet Forgery By ISPs: A

Report On The Comcast Affair, ,November 2007 <https://www.eff.org/files/eff_comcast_report.pdf>.

[13] F. Lehrieder, G. Dán, T. Hoßfeld, S. Oechsner, V. Singeorzan, The Impactof Caching on BitTorrent-like Peer-to-peer Systems, in: IEEEInternational Conference on Peer-to-Peer Computing (IEEE P2P),2010.

[14] T. Karagiannis, P. Rodriguez, K. Papagiannaki, Should internet serviceproviders fear peer-assisted content distribution?, in: ACM InternetMeasurement Conference (IMC’05), 2005.

[15] R. Cuevas, N. Laoutaris, X. Yang, G. Siganos, P. Rodriguez, Deep divinginto bittorrent locality, in: INFOCOM, IEEE, 2010.

[16] H. Xie, Y.R. Yang, A. Krishnamurthy, Y.G. Liu, A. Silberschatz, P4P:Explicit Communications for Cooperative Control Between P2P andNetwork Providers, in: ACM SIGCOMM, 2008.

[17] D.R. Choffnes, F.E. Bustamante, Taming the torrent: a practicalapproach to reducing cross-ISP traffic in peer-to-peer systems,SIGCOMM Computer Communication Review 38 (4) (2008) 363–374.

[18] P. Racz, S. Oechsner, F. Lehrieder, BGP-based Locality Promotion forP2P Applications, in: International Conference on ComputerCommunication Networks (ICCCN), 2010.

[19] S. Le Blond, A. Legout, W. Dabbous, Pushing bittorrent locality to thelimit, Computer Networks 55 (3) (2010) 541–557.

[20] M. Piatek, H.V. Madhyastha, J.P. John, A. Krishnamurthy, T. Anderson,Pitfalls for ISP-friendly P2P design, in: Workshop on Hot Topics inNetworks (HotNets), 2009.

[21] C. Huang, J. Li, K.W. Ross, Peer-Assisted VoD: Making Internet VideoDistribution Cheap, in: International Peer to Peer Symposium(IPTPS), 2007.

[22] O. Abboud, K. Pussep, M. Mueller, A. Kovacevic, R. Steinmetz,Advanced Prefetching and Upload Strategies for P2P Video-on-Demand, in: Workshop on Advanced video streaming techniques forpeer-to-peer networks and social networking, 2010.

[23] J. Pouwelse, P. Garbacki, J. Wang, A. Bakker, J. Yang, A. Iosup, D.Epema, M. Reinders, M. van Steen, H. Sips, Tribler: a social-basedpeer-to-peer system, Concurrency and Computation: Practice andExperience 20 (2008) 127–138.

[24] J.D. Mol, J. Pouwelse, M. Meulpolder, D. Epema, H. Sips, Give-to-Get:An Algorithm for P2P Video-on-Demand, in: Multimedia Computingand Networking (MMCN), 2008.

[25] K. Pussep, O. Abboud, F. Gerlach, R. Steinmetz, T. Strufe, AdaptiveServer Allocation for Peer-assisted VoD, in: IEEE InternationalParallel and Distributed Processing Symposium (IPDPS), 2010.

[26] J. Wang, C. Huang, J. Li, On ISP-friendly Rate Allocation for Peer-assisted VoD, in: ACM International Conference on Multimedia (ACMMultimedia), 2008.

[27] J. Seedorf, S. Niccolini, M. Stiemerling, E. Ferranti, R. Winter, QuantifyingOperational Cost-Savings through ALTO-Guidance for P2P LiveStreaming, in: Workshop on Economic Traffic Management (ETM), 2010.

[28] F. Picconi, L. Massoulie, ISP-friend or Foe? Making P2P LiveStreaming ISP-aware, in: International Conference on DistributedComputing Systems (ICDCS), 2009.

[29] K. Kerpez, Y. Luo, F.J. Effenberger, Bandwidth reduction via localizedpeer-to-peer (P2P) video, International Journal of Digital MultimediaBroadcasting (2010) 1–11.

[30] R. Sommer, A. Feldmann, NetFlow: Information Loss or Win?,in: ACM SIGCOMM Workshop on Internet Measurement (IMW),2002.

[31] K. Cho, K. Fukuda, H. Esaki, A. Kato, Observing Slow CrustalMovement in Residential User Traffic, in: ACM InternationalConference on emerging Networking EXperiments andTechnologies (CoNEXT), 2008.

[32] T. Karagiannis, K. Papagiannaki, M. Faloutsos, BLINC: MultilevelTraffic Classification in the Dark, in: ACM SIGCOMM, 2005.

[33] J. Prins, e-Ha ndbook of Statistical Methods, in: C. Croarkin, P. Tobias(Eds.), 6.3.2.4. EWMA Control Charts, NIST/SEMATECH, 2010. URL<http://www.itl.nist.gov/div898/handbook/pmc/Section3/pmc324.htm>.

[34] Ookla, Net Index. URL <http://www.netindex.com/>.

1130 K. Pussep et al. / Computer Networks 56 (2012) 1118–1130

[35] B. Cheng, L. Stein, H. Jin, Z. Zhang, Towards Cinematic Internet Video-on-Demand, in: ACM European Conference on Computer Systems2008 (EuroSys), 2008.

[36] Y. Huang, T.Z. Fu, D.-M. Chiu, C. Huang, J.C. Lui, Challenges, design andanalysis of a large-scale p2p-vod system, in: ACM SIGCOMM, 2008.

[37] C. Huang, A. Wang, J. Li, K.W. Ross, Understanding Hybrid CDN-P2P:Why Limelight Needs its Own Red Swoosh, in: ACM InternationalWorkshop on Network and Operating Systems Support for DigitalAudio and Video (NOSSDAV), 2008.

[38] N. Basher, A. Mahanti, A. Mahanti, C. Williamson, M. Arlitt, AComparative Analysis of Web and Peer-to-Peer Traffic, in: ACMInternational Conference on World Wide Web (WWW), 2008.

[39] F.T. Johnsen, T. Hafsoe, C. Griwodz, P. Halvorsen, WorkloadCharacterization for News-on-Demand Streaming Services, in: IEEEInternational Performance, Computing, and CommunicationsConference (IPCCC), 2007.

Konstantin Pussep studied computer scienceand received his Master degree from theTechnische Universität Darmstadt, Germanyin 2005. He was working as Research Assis-tant at the Multimedia Communication Lab atTechnische Universität Darmstadt and fin-ished his Phd (Dr.-Ing.) in Electrical Engi-neering & Information Technology in 2011. Inhis work he has dealt with peer-assistedcontent distribution, focusing on the video-on-demand streaming. His research interestsinclude optimized utilization of overlay

resources, network-aware traffic management, and energy-efficiency ofend-user devices.

Frank Lehrieder studied computer scienceand business administration at the Universityof Würzburg/Germany and University Anto-nio de Nebrija in Madrid/Spain. In May 2008,he received his Master degree in computerscience from University of Würzburg. Now, heis a researcher at the Chair of CommunicationNetworks in Würzburg and pursuing his PhD.His special interests are performance evalua-tion and resource management for peer-to-peer networks and Future Internet infra-structures.

Christian Gross studied Computer Science atTechnische Universität Darmstadt and grad-uated as Master of Science in April 2009 withhis thesis ‘‘Development and Evaluation of aMonitoring Framework for P2P Systems’’.Since May 2009 he is a PhD student at theMultimedia Communication Lab (KOM) atTechnische Universität Darmstadt. Hisresearch focuses on benchmarking of peer-to-peer systems and on constructing energy-aware overlays for mobile peer-to-peer sce-nario.

Simon Oechsner studied computer scienceand received his Master degree from theUniversity of Würzburg, Germany in 2004. Hefinished his PhD studies on ‘‘PerformanceChallenges and Optimization Potential ofPeer-to-Peer Overlay Technologies’’ in 2010under Prof. Tran-Gia at the Chair of Commu-nication Networks at the same University. Hisresearch focuses on analytical and simulativeperformance evaluation of overlays. Hisinterests include search and content distri-bution overlays, covering DHTs as well as

currently popular file-sharing and video-streaming networks.

Markus Günther received his Master degreein Computer Science from the TechnischeUniversität Darmstadt. Later he joined theMultimedia Communication Lab of the sameuniversity. In his work he implemented net-work-friendly mechanisms for peer-assistedcontent distribution systems and performeddistributed evaluation on a large-scale cluster.

Sebastian Meyer received his Master degreein computer science from the TechnischeUniversität Darmstadt in 2010. In his studieshe focused on net centric systems and in hismaster thesis he has dealt with incentive-based traffic management for peer-to-peeroverlays. He is currently working as a soft-ware engineer.