Fast Distribution of Replicated Content to Multi- Homed Clients Mohammad Malli Arab Open University, Beirut, Lebanon

Transcription

1 ACEEE Int. J. on Information Tehnology, Vol. 3, No. 2, June 2013 Fast Distribution of Repliated Content to Multi- Homed Clients Mohammad Malli Arab Open University, Beirut, Lebanon Abstrat Clients an potentially have aess to more than one ommuniation network nowadays due to the availability of a wide variety of aess tehnologies. On the other hand, servie repliation has beome a trivial approah in overlay networks to provide a high availability of data and better QoS. In this paper, we onsider suh a multi-homed lient seeking a repliated servie in overlay network (e.g., CDN, peer-topeer). Our aim is to improve the ontent distribution by proposing a new model for being applied at the appliationlevel and in a fully distributed way. Basially, our model proposes to determine the best mirror server that ould be reahed through eah lient s network interfae based on appliation utility funtion. Then, it onsists of downloading the requested ontent from the determined best servers simultaneously through their assoiated interfaes. Eah best server should deliver a speifi estimated range of bytes (i.e., ontent hunk) to an independent TCP soket opened at the lient side for being finally aggregated at the appliationlevel. Our real experiments show that our model is able to onsiderably improve the QoS (e.g., ontent transfer time) pereived by the lient omparing to the traditional ontent distribution tehniques. Index Terms ontent distribution, servie repliation, multihoming. I. INTRODUCTION Servie repliation is a salable solution for the distribution of digital ontent over the Internet. The need for this repliation is aused by the inreasing number of Internet users and by the desire to improve the QoS. Also, it is important for ahieving a high availability of data. Many overlay networks are proposed and installed to realize this repliation: (i) Content Distributed Networks (CDN), where lient requests are forwarded by request rediretors, and where the ontents are stored in mirror servers geographially distributed over the Internet. Many ompanies, like Akamai [1], provide CDNs to ontent providers. (ii) Peer-to-peer networks (e.g., bittorrent [2], where peers behave as lients and servers. On the other hand, one an profit from multihomed lients to ahieve bandwidth aggregation by striping data aross the multiple network interfaes of the lients. In this paper, we address the problem of improving the transfer time pereived by multi-homed lients when requesting digital ontent repliated in the mirror servers of one CDN network (resp. peer-to-peer network) or in multiple ones (i.e., ontent multi-homing [3], [4], [5]). In the following disussion, we onsider a server as being either a server among a set of repliated servers in a CDN or a peer in a peerto-peer network that hold the requested ontent. The best server is the one whih is able to provide the requested servie to the lient with a better QoS than all other servers. Also, we mean by lient a standard lient in the lient/server paradigm, or a peer that requests ontent in a peer-to-peer network; these terms are used in the paper interhangeably. Clearly, the best server varies from one lient to another based on many parameters as the performane on the path onneting the lient to the server through eah network interfae. For enhaning the ontent distribution, many solutions [6], [7], [8], [9] rely on partiular network infrastruture nodes (e.g., load-aware Anyast router, route ontroller, peer oordinator, et.) taking into aount network-speifi onstraints (e.g., traffi engineering onstraints) pereived by the ISP or the overlay network operator to solve the problem as a global optimization problem. While suh approahes ould be of great benefit for traffi engineering purposes, end-systems solutions is able to provide better enhanement to the performane pereived by the lients. Besides, in some senarios, end-systems solutions are able to ahieve a better traffi engineering outome than the ISPs an by themselves as shown in [10]. Moreover, one an avoid the deployment limitations (e.g., network overhead) of the existing solutions by solving the problem at the end-user level in a fully distributed way. On the other hand, although the onept of multipath-apable end systems is interesting to be applied at the transport level [10], [11], [12], [13], [14], there is no protool that simultaneously uses multiple paths has ever been standardized let alone widely deployed to replae the most widely used existing protool TCP. Therefore, we propose a new model to be applied at the appliation-level and in a fully distributed way for improving the QoS pereived by multi-homed end-users. It onsists in lient downloading a repliated ontent from a ertain set of best mirror servers simultaneously through his/her different network onnetions. Firstly, it proposes to determine the best server that ould be ommuniated through eah network interfae based on appliation utility funtion. Then, it onsists of downloading the requested ontent from the determined best servers simultaneously but with different estimated amounts. This must be ahieved by opening a TCP onnetion with eah best server through its assoiated network interfae to download a speifi estimated range of bytes. The size of this range depends on the weight assigned to the best server; funtion of the performane status on its path to the lient s assoiated network interfae. Thus, our model is able to improve the QoS pereived by the lient

2 ACEEE Int. J. on Information Tehnology, Vol. 3, No. 2, June 2013 through ahieving bandwidth aggregation by striping data aross multiple TCP sokets (i.e., one per network interfae) that download the ontent hunks from their assoiated best servers simultaneously. This paper is organized as follows. The next setion elaborates the problem of best server seletion in repliated servie environment. Then, we present, in Setion III, a new model for distributing repliated ontent to multi-homed lient. The experiments that show the performane enhanement provided by this model are presented in Setion IV. Finally, the onlusion is presented in Setion V. II. SERVICE REPLICATION Servie repliation is a salable solution for the distribution of digital ontent over the Internet. The need for this repliation is aused by the inreasing number of Internet users and by the desire to improve the QoS. Also, it is important for ahieving a high availability of the servie. Many overlay networks (e.g., Content Distributed Networks (CDN), and peer-to-peer networks) are proposed and installed to realize this repliation. The first stage of our approah onsists of determining the best server to be ommuniated through eah lient s network interfae. Many poliies have been studied in the literature for best server seletion. The mostly used approahes an be lassified to the following three ategories: Using the DNS (Domain Name System) to get the IP address of the best server. This widely used tehnique is simple: the DNS servers distribute the IP addresses of multiple servers assoiated to a unique name with a round robin algorithm. It is lear that this solution is not designed to improve the QoS sine it does not onsider any stati or dynami performane limitations. It only ensures basi load balaning. Offering the lient a list of servers and let him hoose manually the best server to ontat. The lient hoie in this ase is based on his own riteria, for example the geographial proximity. Choosing the losest server in terms of delay. Inferring the delay loseness between lient and servers an be done using one of the salable approahes presented in the literature [15], [16], [17], [18], [19], [20], [21]. Most of these solutions are based on the network embedding. Suh approahes are based either on network oordinates or on distane matrix fatorization. Also, the loseness an be determined by identifying the bin of the lient and eah server (see [22]). This an be done by measuring their RTT (Round-Trip Time) to a set of landmark points. By knowing the bins of the lient and servers, the DNS server an lassify the servers (from the best one to the worst one) based on the distane between their bins and the lient s one. Thus, most of the existing solutions for best server seletion are based on simple metris suh as the delay, and the geographial loations whih are unorrelated with other network harateristis (e.g., available bandwidth, loss rate) as pereived in the literature [23]. Hene, these metris are not enough to haraterize the proximity given the heterogeneity of the Internet in terms of path harateristis and aess link speed, and the diversity of appliation requirements. We have realized, in [24], that the proximity must be haraterized in a CHESS spae where it is determined at the appliation level taking into onsideration the network metris that deide on the appliation performane. Therefore, we have proposed to do that using a utility funtion that models the quality pereived by peers at the appliation level. In this framework, a peer is loser than another one to some third peer if it provides a better utility funtion, whatever the position of eah peer in the geographial and delay spaes. For example, take the ase where the servie onsists of lients downloading digital ontent from a set of repliated servers using the TCP protool and where the QoS provided to lients is maximized if the transfer time is minimized. In this ase, hoosing the best server amounts to downloading the file from the server that is able to provide the minimum transfer time. This improves the QoS provided to lients and avoids network and server ongestion by distributing the load over servers and network paths that are less loaded than others. While the haraterization of the proximity in CHESS [24] has a good impat on appliation performane, it is a hallenging task due to the two following major requirements. First, it requires the identifiation of the appropriate utility funtion for eah appliation in a first stage. To solve this problem, many interesting models have been proposed in the literature (e.g., transfer time predition [25], speeh quality predition [26], [27]). The seond hallenging task is the measurement of the different network parameters that impat the utility funtion. This is diffiult to ahieve in large sale networks where the number of peers an be huge. In suh ase, the ost of the diret probing among peers may outweigh the profit of the haraterized proximity. Hene, the estimation of the network parameters, impating the utility funtion, must be ahieved in an easy and salable way. In other terms, this should be ahieved with a small measurement overhead and a limited ooperation among nodes. Partiularly, the determination of the network parameters, on the paths joining a large number of peers, must be ahieved in a way that avoids the diret probing among them as have been proposed in the literature [15], [16], [17], [18], [19], [20], [21], [24], [28]. III. ENHANCED MODEL FOR DISTRIBUTING REPLICATED CONTENT A. Proximity Model IN MULTI-HOMING ENVIRONMENT The major ontribution that we present in this paper is an extended model of the previously proposed one CHESS [24] whih has been very briefly desribed in the previous setion. In the new model, we take advantage of the presene of multihoming environment where multiple network onnetions held at the peer side to improve the pereived performane when downloading ontent from a set of mirror servers. In this setting, we propose to onstrut one CHESS spae per

3 ACEEE Int. J. on Information Tehnology, Vol. 3, No. 2, June 2013 network onnetion. Thus, one an onsider the new model as multi-dimensional CHESS where peers are ranked in eah CHESS spae based on the proximity haraterized in this spae; we reall that the proximity haraterization is based on the appliation utility funtion. Then, a peer ould position itself in an overlay network or deide on the size of the ontent portions to download from the best peer of eah CHESS spae based on the proximity haraterization in suh overlay networks; our fous in this paper is on ontent distribution and not on overlay onstrution. Partiularly, downloading ontent from multiple best servers (i.e., pooling apaity over the spae) through the different network onnetions simultaneously (i.e., pooling apaity over the time) ould be ahieved obviously in shorter time delay than that ahieved by the trivial pointto-point ontent delivery tehniques. Besides, this an provide better availability of ontent and resilieny of servie. To the best of our knowledge, there is no model that pools apaity over the spae and the time as effiient as the one proposed in this paper although the onept of resoure pooling has been widely elaborated in the literature. Modern approahes rely on end systems for managing the network traffi patterns and enhaning the ontent distribution. In this sope, many solutions [6], [7], [8], [9] rely on partiular network infrastruture nodes (e.g., load-aware Anyast router, route ontroller, peer oordinator, et.) taking into aount network-speifi onstraints (e.g., traffi engineering onstraints) pereived by the ISP or the overlay network operator to solve the problem as a global optimization problem. While suh approahes ould be of great benefit for traffi engineering purposes, end-systems solutions is able to provide better performane enhanement pereived by the lients. Besides, in some senarios, end-systems solutions are able to ahieve a better traffi engineering outome than the ISPs an by themselves as shown in [10]. Moreover, one an avoid the deployment limitations (e.g., network overhead) of the existing solutions by solving the problem at the enduser level in a fully distributed way. On the other hand, although the onept of multipath-apable end systems is interesting to be applied at the transport level [10], [11], [12], [13], [14], there is no protool that simultaneously uses multiple paths has ever been standardized let alone widely deployed to replae the most widely used existing protool TCP. Therefore, we take advantage of these pereptions to propose a new model for improving the distribution of repliated ontent to multi-homed peer. Basially, it proposes to determine the best server that ould be ommuniated through eah network interfae based on appliation utility funtion (i.e., the losest peer in eah CHESS spae). Then, it onsists of downloading the requested ontent from the different best servers simultaneously but with different estimated amounts. This must be ahieved by opening a TCP onnetion with eah best server through its assoiated network interfae to download a speifi estimated range of bytes that ould be onsidered as a hunk of the requested ontent. The size of suh hunk depends on the weight assigned to the best server; funtion of the performane status on its path with the lient s assoiated network interfae. Thus, our model is able to improve the QoS pereived by the lient through ahieving bandwidth aggregation by striping data aross multiple TCP sokets (i.e., one per network interfae) that download the ontent hunks from their assoiated best servers simultaneously. More formally, suppose that the network ontains n peers p ={p 1,p 2,...,p n } where eah peer ould play the role of a lient seeking a ontent or a server holding the requested ontent. Obviously, the ontent is repliated in multiple peers (resp. mirror servers). The utility funtion (e.g., delay, available bandwidth, predited download time) on the paths joining peers p i and p j (i,j={1...n}) on top of network onnetion (={1...k}) is represented by an n X n matrix U, where U ij is the estimated utility funtion from p i to p j through the network onnetion. The fat that peers ould have different number of network onnetions and thus different values of k does not affet the funtionality of our model sine we are presenting a distributed algorithm to be exeuted at eah peer independently. In ase that every peer has one network onnetion, the system onverges to one CHESS spae where the ontent must be transferred to the lient s unique network interfae from the best server seleted as desribed in Setion II. Thus, for every multi-homed peer p i, the rest of peers p j are ranked through every network onnetion based on the estimations U ij. We assume in this model that the larger the utility funtion value, the better the quality of servie (e.g., available bandwidth on the network path onneting peers) and the loser the peers to eah other in this spae. Obviously, in ase where the utility funtion is in ontrast signifiant for small values (e.g., delay, predited download time), peers must be ranked aording to the inreasing order of U ij. its losest peer in the CHESS spae (i.e., best peer reahable through network onnetion ) is the peer P i that satisfies MAX j={1..n} U ij (resp. MIN j={1..n} U ij in ase the utility funtion is signifiant for small values). The best peer P i must be different than the ones determined through the other network interfaes even if it is the losest peer to p i in the different CHESS spaes. Thus, if the losest peer in the CHESS spae is the same one seleted as the best peer through another network onnetion, then P i must be seleted as the next loser peer (based on the previously presented ranking) that is not yet seleted as the best peer through another network onnetion. Hene, eah best server an upload only one ontent hunk to a peer through one of its network onnetion. In this way, we are able to pool the apaity over the spae by relying on a good number of best servers (i.e., equal to the number of lient s network onnetions) instead of uploading the hunks from a fewer number of best servers (i.e., smaller than the number of lient s

4 ACEEE Int. J. on Information Tehnology, Vol. 3, No. 2, June 2013 network onnetions). Besides, the ontent must be delivered from the seleted best servers in a onurrent way to pool the apaity over the time as well. Moreover, for more effiient pooling of the servers apaity, the size of a ontent hunk to be downloaded from eah best server must be proportional to the performane on its end-to-end path with the assoiated lient s network onnetion (as will be evaluated in this model). its best CHESS spae is the spae C that satisfies max U. Then, the losest peer to p in the best CHESS ={1..k} i(pi ) i C spae C is denoted by P ic. Thus, P i is loser to p i than the other best peers of the rest CHESS spaes sine the C estimated utility funtion between peer p i and P i has the greatest value. Our algorithm an be useful for improving the overlay onstrution and ontent distribution. For overlay onstrution, a peer ould infer its losest peer in eah CHESS spae to better position itself in an overlay network; this issue will be explored and examined in a future work. As for ontent delivery purpose whih is our fous in this paper, it onsists of determining the orrespondent proportion to download DP i by peer p i through its network onnetion from its assoiated best server. Suh proportion depends on the relative value of the performane on the path onneting p i to P i through the network onnetion with respet to those on its path with the other best servers reahable through the remaining network onnetions. Therefore, we propose to evaluate DP i as the fration between the weight w i assigned to P i and the overall weights assigned to all the seleted best servers: where, and the weight w i evaluates how good is the predited quality between p i and P i omparing to the best ase estimated between p i and P ic. Therefore, one an evaluate it as the fration of the utility funtion evaluated between p i and the best peer in the CHESS spae (i.e., to be ommuniated (1) (2) Thus, we propose that peer p i opens a TCP onnetion with the best server seleted in eah CHESS spae P i (={1...k}) to download the ontent hunk having the following proportion from the whole ontent: If these proportions do not divide the file size into finite ranges, the residual value is added to the range of bytes alloated to the losest peer in the best CHESS spae P ic. Finally, the reeiving appliation at the lient side does the re-sequening using a buffer having the size of the requested ontent to ahieve reliable in-sequene data deliv.ery. B. Case study Take the example of Figure 1 where peer p 1 would like to download a ontent of 999B repliated in peers p 2, p 3, and p 4. In this senario, we assume that there are two network onnetions 1 and 2 that ould be used by peers to ommuniate with eah other. Thus, there are two orrespondent CHESS spaes. Besides, the utility funtion is assumed to be a simple metri whih is the available bandwidth on the end-to-end network path for simpliity seeking. As shown in the figure, the available bandwidth values on the paths onneting p 1 to the other peers through network onnetion 1 are: p 1 to p 2 is 50 Mbps. So, U 12 1 = 50, p 1 to p 3 is 25 Mbps. So, U 13 1 = 25, p 1 to p 4 is 10 Mbps. So, U 14 1 = 10. For the seond CHESS spae whih is reahable through the network onnetion 2, the available bandwidth values are: 2 p 1 to p 2 is 25 Mbps. So, U 12 = 25, 2 p 1 to p 3 is 50 Mbps. So, U 13 = 50, 2 p 1 to p 4 is 100 Mbps. So, U 14 = 100. In this ase, p 2 is the best peer for p 1 in CHESS spae 1 with U 12 1 = 50 and p 4 is its best peer in CHESS spae 2 with U 14 2 = 100. Thus, the best CHESS spae for p 1 is 2 (i.e., C= 2 ). Then, peer p 1 assigns the following weights to the best servers determined in the two CHESS spaes 1 and 2 respetively: (3) (4) Figure 1. An example of multi-homed peer seeking a repliated ontent through the network interfae ) over the utility funtion between p i and its losest peer in the best CHESS spae: and, (5) (6)

5 ACEEE Int. J. on Information Tehnology, Vol. 3, No. 2, June 2013 Thus, peer p 1 should download from peer p 2 through the network interfae of CHESS spae 1 a ontent hunk having the following proportion: and subsequently the range of bytes [1,333] from the whole ontent to download. Conurrently, peer p 1 should download from peer p 4 through the interfae of CHESS spae 2 a ontent hunk having the following proportion: and subsequently the range of bytes [334,999] from the whole ontent to download. If the file size is 1000B whih ould not be divided to a finite ranges of bytes in this ase. Then, the residual extra byte from the division is alloated to the range alloated to the losest peer p 4 in the best CHESS spae 2. Then, the range of bytes to be downloaded from p 4 beomes [334, 1000] and the range of bytes to be downloaded from p 2 remains the same. IV. ENHANCED CONTENT DELIVERY PERCEIVED BY THE APPLICATION For evaluating the improvement that an be ahieved by applying the presented approah for ontent distribution, we have onduted real experiments having the following settings. We take 10 multi-homed lients spread in the (7) (8) T m whih is the lateny measured when applying our new model. In this ase, the ontent is delivered onurrently from the two best servers identified through the two network onnetions aording to the multidimensional CHESS model presented in Setion III. T b whih is the lateny measured when downloading the ontent in point-to-point mode from the best server identified through the DSL onnetion aording to the CHESS model published in [24] and briefly presented in Setion II. The two ways of ontent delivery applied in eah trae have been ahieved in a sequential manner. The whole traes have taken plae in different dates and times during the month of July We assume in these experiments that the utility funtion, for best server seletion, is the end-to-end available bandwidth on the network path onneting the lient to the server. One an rely on the predited transfer time metri [25] as a more optimized tool in this sope. However, this hoie of utility funtion does not affet the observations of our results sine our aim is to ompare our new model of ontent delivery with the lassi way of point-to-point ontent delivery despite the way of determining the best servers. Hene, to measure the available bandwidth, we have applied the probe rate model [29], [30]. This is ahieved by sending a stream of pakets from the lient to the server at a rate greater than 512kbps sine the end-to-end available bandwidth is surely smaller (or equal) than this value whih is the maximum download rate of the lient per network interfae. Then, the end-to-end available bandwidth is Figure 2. Enhanement of the ontent delivery time networks of four ISPs where eah lient is plugged simultaneously to two network onnetions whih are DSL and WiMax. The DSL and WiMax onnetions have the same dediated bandwidth of 512kbps download speed and 128kbps upload speed. Eah trae onsists in multi-homed lient downloading a digital ontent of partiular size from one of 20 seleted CDN networks having worldwide distributed sets of mirrors. This leads to a total of 200 traes where in eah trae we measure the following two ontent delivery times: Figure 3. Enhanement variation with respet to the file size alulated as the rate of the reeiving stream s ehoes. Thus, to ompare our new ontent delivery model with the point-to-point one, we evaluate the enhanement as a metri having the following expression: In Figure 2, we plot the CDF (i.e., Cumulative Distribution Funtion) of the metri enhanement estimated from the 200 traes. The figure shows that for around 85% of the traes, our new model for ontent distribution improves the delivery time by a proportion between 0.4 and 0.9. The rest of traes (9)

6 ACEEE Int. J. on Information Tehnology, Vol. 3, No. 2, June 2013 (i.e., around 15%) shows an enhanement between 0.1 and 0.4. The onsiderable enhanement of ontent delivery time an be also observed from the expeted value of the enhanement metri (averaged over the whole traes) whih is obtained equal to In other terms, our traes show that with high probability, the delivery time ould be dereased to an amount smaller than roughly half of its value when applying our model for ontent delivery instead of the lassi pointto-point delivery method. To study the variation of the enhanement with respet to the file size, we plot it in Figure 3 where the first metri is on the y-axis and the seond one is on x-axis. Every point in the graph represents the average enhanement of several traes transferring ontents of partiular size in megabytes. One an observe from the figure that while the enhanement flutuates around the value 0.5, the graph shows a positive orrelation between the enhanement metri and the file size. Moreover, other measures show that when the file size inreases onsiderably over 50MB till the value 250MB, the enhanement flutuates between 0.6 and 0.7. This interval of file sizes has not been plotted in Figure 3 sine its traes ontain only one trae every inrease of 10MB; thus, we are unable to plot the average enhanement values over this interval. However, a positive orrelation has been also learly pereived through the oeffiient whih obtained equal to Hene, one an realize that our model for ontent delivery is able to derease more onsiderably the lateny when transferring larger digital ontents by striping the data aross multiple TCP sokets (i.e., one per network interfae). This observation an be due to the fat that when the ontent size inreases the ongestion avoidane phase beomes more dominant than the slow start phase of the download onnetion; we notie that the ongestion window size inreases exponentially in the slow start phase and then linearly in the ongestion avoidane phase. In this ase, it is obvious to observe better enhanement when reduing greater number of rounds in the ongestion avoidane phase and spending more rounds in the slow start phase. V. CONCLUSIONS AND PERSPECTIVES In this paper, we propose a new model to improve the ontent distribution in overlay networks. Our model takes advantage from the ontent repliation and multi-homing failities whih are widely available nowadays. This is done by determining the best peer that ould be reahed through eah network interfae based on the estimation of the appliation utility funtion. Then, it onsists of ahieving bandwidth aggregation by striping data aross multiple TCP sokets (i.e., one per network interfae) that download the ontent hunks from their assoiated best servers simultaneously. Our extensive real measurements show learly how our solution outperforms the existing solutions by dereasing onsiderably the ontent distribution time. Our traes show that with high probability, the delivery time ould be dereased to an amount smaller than roughly half of its value when applying our model for ontent delivery instead of the lassi point-to-point ontent delivery methods. This is due to the fat that it ombines the best server seletion sheme with the bandwidth aggregation faility in multi-homing environment. Also, the results show that it is able to provide better quality of servie when distributing larger ontent. Besides, it is more flexible to be deployed than the solutions that depend on network infrastruture nodes and those proposed at the transport layer. Thus, our approah does not require the deployment of any speial network node and does not impose any hange to the existing reliable transport protool TCP. This is ensured by working at the appliation-level and in a fully distributed way. Conerning the hallenge of determining the best server, many solutions have been proposed for estimating the appliation utility funtion in a salable way. This ould be done by firstly defining a funtion of the parameters impating appliation performane and then relying on a salable approah for inferring these parameters using a limited set of measurements. Regarding our future work, we will test the effiieny of the presented model when used for overlay onstrution. Besides, we will investigate how our approah reats to ongestion. ACKNOWLEDGMENT The author wishes to thank Mr. Ahmad Moosa for his valuable help in the testing phase. REFERENCES [1] Akamai, [2] BitTorrent, [3] H. H. Liu, Y. Wang, Y. R. Yang, H. Wang, and C. Tian, Optimizing ost and performane for ontent multihoming, ACM SIGCOMM, [4] V. K. Adhikari, Y. Guo, F. Hao, M. Varvello, V. Hilt, M. Steiner, and Z.-L. Zhang, Unreeling netflix: Understanding and improving multi-cdn movie delivery, IEEE INFOCOM, [5] G. Bertrand, E. Stephan, G. Watson, T. Burbridge, P. Eardley, and K. Ma, Use ases for CDNi, IETF Draft, [6] H. A. Alzoubi, S. Lee, M. Rabinovih, O. Spatshek, and J. Van Der Merwe, A pratial arhiteture for an anyast CDN, ACM Transations on the Web, vol. 5, pp , [7] R. S. Peterson, B. Wong, and E. G. Sirer, A ontent propagation metri for effiient ontent distribution, ACM SIGCOMM, [8] I. Poese, B. Frank, B. Ager, G. Smaragdakis, and A. Feldmann, Improving ontent delivery using provider-aided distane information, ACM IMC, [9] R. S. Peterson and E. G. Sirer. Antfarm, effiient ontent distribution with managed swarms, NSDI, [10] D. Wishik, M. Handley, M. Braun, The resoure pooling priniple, ACM SIGCOMM CCR, vol. 38, pp , [11] D. Wishik, M. Handley and C. Raiiu, Control of multipath TCP and optimization of multipath routing in the Internet, NetCOOP, [12] M. Zhang, J. Lai, A. Krishnamurthy, L. Peterson, and R. Wang, A Transport Layer Approah for Improving End-to-End

7 ACEEE Int. J. on Information Tehnology, Vol. 3, No. 2, June 2013 Performane and Robustness Using Redundant Paths, USENIX, [13] H. Hsieh and R. Sivakumar, ptcp: An End-to-End Transport Layer Protool for Striped Connetions, IEEE ICNP, Paris, [14] H. Hsieh and R. Sivakumar, A Transport Layer Approah for Ahieving Aggregate Bandwidths on Multi-homed Mobile Hosts, ACM MOBI-COM, Atlanta, GA USA, [15] E. Ng and H. Zhang, Prediting Internet network distane with oordinates-based approahes, IEEE Infoom, [16] L. Tang, and M. Crovella, Virtual Landmarks for the Internet, ACM IMC, [17] B. Wong, A. Silvkins, and E. G. Sirer, Meridian: A Lightweight Network Loation Servie without Virtual Coordinates, ACM SIGCOMM, [18] P. Franis, S. Jamin, C. Jin, Y. Jin, D. Raz, Y. Shavitt, and L.Zhang, IDMaps: A Global Internet Host Distane Estimation Servie, IEEE/ACM Transations on Networking, [19] Y. Shavitt, and T. Tankel, Big-Bang Simulation for Embedding Network Distanes in Eulidean Spae, IEEE Infoom, [20] H. Lim, J. Hou, and C. Choi. Construting Internet Coordinate System based on Delay Measurement, ACM IMC, [21] J. Ledlie, P. Gardner and M. Seltzer, Network Coordinates in the Wild, NSDI, [22] S. Ratnasamy and M. Handly, R. Karp and S. Shenker, Topologially-Aware Overlay Constrution and Server Seletion, IEEE Infoom, [23] M. Malli, C. Barakat, and W. Dabbous, Appliation-level versus Network-level Proximity, Asian Internet Engineering Conferene, Thailand, [24] M. Malli, C. Barakat, and W. Dabbous, CHESS: An Appliation-aware Spae for Enhaned Salable Servies in Overlay Networks, IEEE Computer Communiation Journal, vol. 31, pp , [25] M. Malli, C. Barakat, and W. Dabbous, An Effiient Approah for Content Delivery in Overlay Networks, IEEE CCNC, [26] L. Ding, and R. Goubran, Speeh Quality Predition in VoIP Using the Extended E-Model, IEEE Globeom, [27] R. Cole, and J. Rosenbluth, Voie over IP Performane Monitoring, ACM SIGCOMM CCR, vol. 31, pp. 9-24, [28] N. Hu, P. Steenkiste, Exploiting Internet Sharing for Large Sale Available Bandwidth Estimation, ACM IMC, [29] N. Hu and P. Steenkiste, Evaluation and Charaterization of Available Bandwidth Tehniques, IEEE JSAC Speial Issue in Internet and WWW Measurement, Mapping, and Modeling, [30] V. J. Ribeiro, R. H. Riedi, R. G. Baraniuk, J. Navratil, and L. Cottrell, PathChirp: Effiient Available Bandwidth Estimation for Network Paths, Passive and Ative Measurement Workshop, 2003.