An experimental study of the skype peer to-peer vo ip system


Published on

An experimental study of the skype peer to-peer vo ip system

Published in: Technology, Business
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

An experimental study of the skype peer to-peer vo ip system

  1. 1. An Experimental Study of the Skype Peer-to-Peer VoIP System Saikat Guha Neil Daswani, Ravi Jain Cornell University Google {daswani, ravijain}@google.comABSTRACT sharing usage? Gummadi et al. [13], find that file-sharingDespite its popularity, relatively little is known about the traffic users are patient when their file-searches succeed and leavecharacteristics of the Skype VoIP system and how they differ from their client online for days until their requests are completed,other P2P systems. We describe an experimental study of Skype and close their clients within minutes when searches fail. InVoIP traffic conducted over a five month period, where over 82 mil- contrast, we find that Skype users regularly run the clientlion datapoints were collected regarding the population of online during normal working hours, and close it in the evening,clients, the number of supernodes, and their traffic characteristics. leading to different network dynamics. We also considerThis data was collected from September 1, 2005 to January 14, the overall utilization and resource consumption of Skype.2006. Experiments on this data were done in a black-box manner, Does Skype really need the resources of millions of peers toi.e., without knowing the internals or specifics of the Skype system provide a global VoIP service, or can a global VoIP service beor messages, as Skype encrypts all user traffic and signaling traffic supported by a limited amount of dedicated infrastructure?payloads. The results indicate that although the structure of the We find that the median network utilization in Skype peersSkype system appears to be similar to other P2P systems, particu- is very low, but that peak usage can be high.larly KaZaA, there are several significant differences in traffic. The Overall, our work makes three contributions. First, innumber of active clients shows diurnal and work-week behavior, §2, we shed light on some design choices in the proprietarycorrelating with normal working hours regardless of geography. Skype network and how they affect robustness and avail-The population of supernodes in the system tends to be relatively ability. Second, in §3 and §4 respectively, we analyze nodestable; thus node churn, a significant concern in other systems, dynamics and churn in Skype’s peer-to-peer overlay, and theseems less problematic in Skype. The typical bandwidth load on asupernode is relatively low, even if the supernode is relaying VoIP network workload generated by Skype users. Third, we pro-traffic. vide data on user-behavior that can be used for future design The paper aims to aid further understanding of a significant, and modeling of peer-to-peer VoIP networks; note that de-successful P2P VoIP system, as well as provide experimental data veloping an explicit quantitative model is out of scope ofthat may be useful for future design and modeling of such sys- the present paper. Altogether, we find evidence that Skypetems. These results also imply that the nature of a VoIP P2P system is fundamentally different from the peer-to-peer networkslike Skype differs fundamentally from earlier P2P systems that are studied in the past.oriented toward file-sharing, and music and video download appli-cations, and deserves more attention from the research community. 2. SKYPE OVERVIEW Skype offers three services: VoIP allows two Skype users to1. INTRODUCTION establish two-way audio streams with each other and supports conferences of up to 4 users, IM allows two or more Skype Email was the original killer application for the Internet. users to exchange small text messages in real-time, and file-Today, voice over IP (VoIP) and instant messaging (IM) are transfer allows a Skype user to send a file to another Skypefast supplementing email in both enterprise and home net- user (if the recipient agrees)1 . Skype also offers paid servicesworks. Skype is an application that provides these VoIP and that allow Skype users to initiate and receive calls via regularIM services in an easy-to-use package that works behind telephone numbers through VoIP-PSTN gateways.Network Address Translators (NAT) and firewalls. It has at- Despite its popularity, little is known about Skype’s en-tracted a user-base of 50 million users, and is considered valu- crypted protocols and proprietary network. Garfinkel [11],able enough that eBay recently acquired it for more than $2.6 concludes that Skype is related to KaZaA; both the compa-billion [18]. In this paper, we present a measurement study of nies were founded by the same individuals, there is an overlapthe Skype P2P VoIP network and obtain significant amounts of technical staff, and that much of the technology in Skypeof data. This data was collected from September 1, 2005 to was originally developed for KaZaA. Network packet levelJanuary 14, 2006 at Cornell University. While measurement analysis of KaZaA [16] and of Skype [1] support this claimstudies of both P2P file-sharing networks [26, 27, 2, 13, 21] by uncovering striking similarities in their connection setup,and “traditional” VoIP systems [14, 17, 4] have been per- and their use of a “supernode”-based hierarchical peer-to-formed in the past, little is known about VoIP systems that peer network.are built using a P2P architecture. Supernode-based peer-to-peer networks organize partic- One of our key goals in this paper is to understand how P2P ipants into two layers: supernodes, and ordinary nodes.VoIP traffic in Skype differs from traffic in P2P file-sharing Such networks have been the subject of recent researchnetworks and from traffic in traditional voice-communication in [29, 28, 6, 5]. Typically, supernodes maintain an overlaynetworks. For example, how does the interactive-nature ofVoIP traffic affect node session time (and thus complicate 1 This is different from file-sharing in Gnutella, KaZaA and BitTor-overlay maintenance) as compared to the non-interactive file- rent, where users request files that have been previously published.
  2. 2. network among themselves, while ordinary nodes pick one VoIP and file-transfer sessions.(or a small number of) supernodes to associate with; supern- Expt. 4: Supernode and client population. In this experiment,odes also function as ordinary nodes and are elected from we discovered IP addresses and port numbers of supernodesamongst them based on some criteria. Ordinary nodes issue between Jul. 25, 2005 and Oct. 12, 2005. Each client cachesqueries through the supernode(s) they are associated with. a list of supernodes that it is aware of. We wrote a script Expt. 1: Basic operation. We conducted an initial experi- that parses the Skype client’s supernode-cache and adds thement to examine the basic operation and design of the Skype addresses in the cache to a list. Our script then replacesnetwork in some more detail. We ran two Skype clients the cache with a single supernode address from the list such(version for Linux) on separate hosts, and observed that the client is forced to pick that supernode the next timethe destination and source IP addresses for packets sent and the client is run. The script starts the client and waits forreceived in response to various application-level tasks. We it to download a fresh set of supernode addresses from theobserved that in Skype, ordinary nodes send control traffic supernode to which it connects. The script then kills theincluding availability information, instant messages, and re- client causing it to flush its supernode-cache. The cache isquests for VoIP and file-transfer sessions over the supernode processed again and the entire process repeated; the result ispeer-to-peer network. If the VoIP or file-transfer request a crawl of the supernode network which discovers supern-is accepted, the Skype clients establish a direct connection ode addresses. Our experiment discovered 250K supernodebetween each other. To examine this further, we repeated addresses, and was able to crawl 150K of them. As a side-the experiment for a single client behind a NAT2 , and both effect, the script also records the number of online Skypeclients behind different NATs. We observed that if one client users each time the client is run, as reported by the Skypeis behind a NAT, Skype uses connection reversal whereby the client.node behind the NAT initiates the TCP/UDP media session Expt. 5: Supernode presence. In this experiment, we gath-regardless of which end requested the VoIP or file-transfer ered “snapshots” of which supernodes were online at a givensession. If both clients are behind NATs, Skype uses STUN- time. We wrote a tool that sends application-level pings tolike NAT traversal [25, 10] to establish the direct connection. supernodes; the tool replays the first packet sent by a SkypeIn the event that the direct connection fails, Skype falls back client to a supernode in its cache, and waits for an expectedto a TURN-like [24] approach where the media session is response. For each snapshot, we perform parallel pings torelayed by a publicly reachable supernode. This latter ap- a fixed set of 6000 nodes randomly selected from the set ofproach is invoked when NAT traversal fails, or a firewall supernodes discovered in the second experiment. Each snap-blocks some Skype packets. Thus the overall mechanism shot takes 4 minutes to execute. These snapshots are taken atthat Skype employs to service VoIP and file transfer requests 30 minute intervals for one month beginning Sep. 12, quite robust to NAT and firewall reachability limitations. Skype encrypts all TCP and UDP payloads, therefore, our Expt. 2: Promotion to supernode. We next investigated how analysis on this dataset is restricted to IP and TCP/UDPnodes are promoted to supernodes. In an experiment we con- header fields including the source and destination IP addressducted, we ran several Skype nodes in various environments and port, and packet lengths.and waited two weeks for them to become supernodes. ASkype node behind a saturated network uplink, and one be- 4. CHARACTERIZING SKYPE’S NETWORKhind a NAT, did not become supernodes, while a fresh install Churn in P2P networks, the continuous process of nodeson a public host with a 10 Mbps connection to the Internet joining and leaving the system, increases routing latency asjoined the supernode network within minutes. Consequently, some overlay traffic is routed through failed nodes, whileit appears that Skype supernodes are chosen from nodes that some new nodes are not taken advantage of. Many peer-have plenty of spare bandwidth, and are publicly reachable. to-peer networks handle churn by dynamically restructur-This approach clearly favors the overall availability of the ing the network through periodic or reactive maintenancesystem. We did not test additional criteria such as a history traffic. Churn has been studied extensively in peer-to-peerof long session times, or low processing load as suggested file-sharing networks [26, 27, 2, 13, 7]; the consensus is thatin [28]. As we show later, the population of supernodes se- churn can be high. The session time of a node is defined aslected by Skype, apparently based on reachability and spare the time between when a node joins the network, and thenbandwidth, tends to be relatively stable. Skype, therefore, subsequently leaves. It has been found for other P2P net-represents an interesting point in the P2P design-space where works that mean node session time can be as low as a fewheterogeneity is leveraged to control churn, not just cope with minutes, and that frequent updates are needed to maintainit. consistency [23]. In this section, we study churn in Skype’s supernode P2P3. METHODOLOGY network, using the data derived from Expts. 3-5 described In order to understand the Skype network, we performed above. (Note that we are focusing on the supernode net-three experiments in parallel. work, not the ordinary node network, in this study). We find Expt. 3: Supernode network activity. In this experiment, we that there is very little churn in the supernode network, andobserved the network activity of a Skype supernode for 135 that supernodes demonstrate diurnal behavior causing mediandays (Sep. 1, 2005 to Jan. 14, 2006). We ran version session times of several hours. Further, we find that sessionof the Skype binary for Linux (packaged for Fedora Core 3), lengths are heavy-tailed and are not exponentially distributed.and used ethereal [8] to capture the 13GB of data sent and Figure 1(a) shows that the number of Skype supernodesreceived by the supernode during this time, including relayed is more stable than the number of online Skype users. The figure is split into two parts for clarity; the plot on the left2 We overload NAT to mean NATs and firewalls in the rest of the tracks daily variations in client and supernode populationspaper. from Sep. 18 to Oct. 4, while the plot on the right zooms-in on
  3. 3. 10% Joins 100% Departs 90% 5% 80% Supernode Churn 70% Fraction Online 60% 0% 50% 40% -5% 30% 20% 10% All Nodes -10% Supernodes Sun 9/18 Sun 9/25 Sun 10/2 0h 6h 12h18h24h 0% Sun 9/18 Sun 9/25 Sun 10/2 0h 6h 12h18h24h Date (noon UTC) UTC Thu 9/22 Date (noon UTC) UTC Thu 9/22 Figure 2: Fraction of supernodes joining or departing the network (a) over the duration of our trace. 1 Supernodes 100% North America Others 80% 0.1 P[session time > x] Distribution of Supernodes Asia 60% 0.01 40% Europe 20% 0.001 1h 2h 4h 8h 12h 1d 2d 4d 8d 16d Session Time 0% Sun 9/18 Sun 9/25 Sun 10/2 0h 6h 12h18h24h Date (noon UTC) UTC Thu 9/22 Figure 3: Log-log plot of the complimentary CDF of supernode session times. (b) pernodes, with its contribution peaking around 11am UTC (mid-day over most of Europe). North America contributesFigure 1: (a) Percentage of all nodes, and supernodes active at any 15–25% of supernodes, with a peak contribution around noontime. (b) Geographic distribution of supernodes as observed over CST. Similarly, Asia contributes 20–25% and peaks aroundthe duration of our trace. its mid-day. Combined with the lower weekend-usage from the previous graph, there is evidence to conclude that Skypehourly variations on Sep. 22. As mentioned in Section 3, the usage, at least for those nodes that become supernodes, isnumber of online users is reported by the official Skype client, correlated with normal working hours. This is differentwhile online supernodes are determined through periodic from P2P file-sharing networks where users download files inapplication-level pings. batches that are processed over days, sometimes weeks [13]. It is clear that there are large diurnal variations with peak Session times reflect this correlation with normal workingusage during normal working hours and significantly reduced hours. As has been observed widely for interactive appli-usage (40–50%) at night. In addition, there are weekly vari- cations like telnet, web, and email [20, 9], node arrivals inations with 20% fewer users online on weekends than on Skype are concentrated towards the morning, while depar-weekdays. The maximum number of users online was 3.9 tures are concentrated towards the evening (Figure 2). Fig-million on Wednesday, Sep. 28 around 11am EST. In com- ure 2 plots the fraction of supernodes joining and leavingparison, of the 6000 randomly-selected supernodes pinged, the network in consecutive snapshots taken at 30 minute in-only 2078 responded to pings at least once during our trace, tervals. The median supernode session time from the sameand between 30–40% of them are online at any given time. experiment is 5.5 hours, as shown in Figure 3. The medianWhile client population varies by over 40% on any given is higher than reported in previous studies of file-sharingday, supernode population is more stable and varies by under networks [26, 27, 2, 13, 7]; however, these studies measure25%. session times for all nodes and not just the supernodes that Figure 1(b) confirms that Skype usage peaks during normal form the P2P overlay. We plot the complementary CDF inworking hours. The graph plots the geographic distribution Figure 3 as it is useful for detecting power law relationships.of active supernodes. Europe accounts for 45–60% of su- We observe that while the supernode session time comple-
  4. 4. mentary CDF is not a strict straight line, i.e., a strict power 1law, it may be approximated as such for our data. 0.9 The nature of this distribution suggests that both arrivals 0.8and departures in Skype cannot be modelled as a Poissonor uniform processes. Consequently, results from past work 0.7that models churn as fixed-rate Poisson processes [23, 5, 15] 0.6may be misleading if directly applied to Skype. CDF 0.5 One way to model churn in Skype is to model node arrival 0.4as a Poisson process with varying hourly rates (higher inthe morning), and picking session times from a heavy-tailed 0.3distribution, similar to the approach used in in [22]. Never- 0.2theless, since there is little churn in the first place (more than 0.1 Control and IM traffic95% of supernodes persist from one thirty-minute snapshot All traffic Relayed VoIP/Filetransferto the next), we expect periodic updates with an update-rate 0 30 100 300 1k 3k 10k 30k 100kchosen accordingly to perform well [23]. As an optimiza- Bandwidth (bps)tion, reactive updates can be used to conserve bandwidth atnight, when there is little churn. Figure 4: Semi-log plot of CDF of bandwidth used by our supernode.5. VOIP IN SKYPE: BANDWIDTH CONSUMP- we resort to statistical approaches for this, which may mis- TION AND PRELIMINARY OBSERVA- classify small file-transfers as control traffic. We separate control traffic from data traffic by identifying the respective TIONS connections as explained in [1]. We further classify the data Skype uses spare network and computing resources of hun- traffic as VoIP or file-transfer based on the ratio of bytes sentdreds of thousands of supernodes, and little additional infras- in the two directions. We use empirical results to conserva-tructure to handle calls, as compared to traditional telephone tively classify data sessions with ∼33 packets transferred percompanies and wireless carriers who rely on expensive, ded- second and an overall one-way bytes-sent ratio greater thanicated, circuit-switched infrastructure. In this section, we 0.2 as VoIP.analyze the role this peer-to-peer network plays in the con- The supernode is engaged in relaying data 9.6% of thetext of VoIP. We find that Skype supernodes incur a small time. This value is smaller than we expected; it is explainednetwork cost for participating in the Skype network. by Skype’s use of NAT traversal that successfully establishes Supernode bandwidth consumption. Figure 4 shows that our direct VoIP/file-transfer sessions through many NATs. ForSkype supernode uses very little bandwidth most of the time. relayed data, the supernode uses 60 kbps in the median (Fig-The bandwidth used by our supernode is plotted for 30 second ure 4).intervals. Fifty-percent of the time, our supernode consumes Sen et al. [27] find that half the users of a P2P file-sharingless than 205 bps. network have upstream bandwidth greater than 56 kbps; Skype VoIP silence suppression. We also observe that Skype does can take advantage of such nodes, when publicly reachable,not use silence suppression and sources 33 packets per sec- to act as supernodes. In addition to the network bandwidth,ond for all VoIP connections regardless of speech charac- our supernode consumed negligible additional processingteristics. Clearly using this simple technique could reduce power, memory and storage as compared to an ordinary node.client bandwidth consumption. It would also reduce supern- VoIP session arrival behavior. Figure 5 offers some insightsode bandwidth because calls that cannot be completed in a into Skype user behavior. These results are preliminary: first,direct peer-to-peer fashion are relayed via a supernode. encryption prevents us from looking into all control traffic, In addition, Skype’s use of peer-to-peer potentially rep- and we therefore are limited to analyzing user behavior onlyresents a convenient looking-glass into a global VoIP/IM for relayed VoIP/file-transfer sessions and not IM or directnetwork. However, this study is limited by the Skype’s en- sessions. This potentially introduces an unavoidable biascryption of control and data messages. The data we have in our user population. Second, this causes us to miss anused for studying VoIP in Skype is extracted from the dataset estimated 85% of the data (based on [12]), resulting in onlyobtained by Expt. 3-5 described previously. However, we 1121 data points for 135 days. Nonetheless, we believeare only able to examine sessions that involve supernodes, that even these preliminary results show interesting trendsand also must make a heuristic estimate of which sessions and, to the best of our knowledge, represent the first publiclyare VoIP sessions and which are file transfer sessions, as we available measurements of call parameters in a VoIP network.describe below. Figure 5(a) suggests that inter-arrival time of relayed VoIP Within these limitations, our observations seem to indicate sessions and file-transfer sessions may be Poisson. File-that Skype calls last longer than calls in traditional telephone transfers are initiated less frequently than voice calls; note,networks, and that files transferred are smaller than in file- however, that file-transfer in this context refers to a usersharing networks. However, we hasten to add that while sending a file to another user (much like email attachments),plausible reasons may exist for such behavior, a larger study and is different from required to validate these observations. In particular, it VoIP session length behavior. Figure 5(b) shows that theis possible that calls that are relayed by supernodes have median Skype call lasted 2m 50s, while the average wasdifferent characteristics than direct peer-to-peer calls. 12m 53s. The longest relayed call lasted for 3h 26m. The Supernode traffic. We separate out low-bandwidth control average call duration is much higher than the 3-minute aver-and IM traffic, and high-bandwidth, relayed VoIP and file- age for traditional telephone calls [19].transfer traffic; due to Skype’s use of encryption, however, One reason for this difference may be that Skype-to-Skype
  5. 5. 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 P[inter-arrival time > x] 0.7 0.6 0.6 0.6 CDF CDF 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.1 0.2 0.1 Relayed VoIP Relayed Filetransfer Relayed Conversation Duration Relayed File Size 0 0.1 0 1m 3m 10m 30m 1h 3h 10h 30h 20s 1m 3m 10m 30m 1h 3h 1k 3k 10k 30k 100k 300k 1M 3M 10M Inter-arrival Time Conversation Duration File Size (bytes) (a) (b) (c)Figure 5: (a) Semi-log plot of CCDF of inter-arrival time of relayed VoIP and file-transfer sessions. (b) Semi-log plot of CDF of Skype VoIPconversation durations. (c) Semi-log plot of CDF of file-transfer sizes.VoIP is free, while phone calls are charged. Bichler and liminary in nature. In particular, further study of the exper-Clarke [3] found that fraudulent pay-phone users, who had imental data results in some preliminary observations thatacquired dialing codes to make long-distance calls for free, indicate that it is possible that Skype calls are significantlywould place calls that lasted 9 minutes on average, while longer than calls in traditional telephone networks, whilelegitimate calls lasted 4 minutes. files transferred over Skype are significantly smaller than The median file-transfer size is 346 kB (Figure 5(c)). The those over file-sharing networks; further work is required tosize is similar to documents, presentations and photos, and validate these preliminary much smaller than audio or video files in file-sharing net- Overall, we present measurement data useful for designingworks [26]. and modeling a peer-to-peer VoIP system. Even though this Altogether, we find that Skype users appear to behave data is limited due to the proprietary nature of Skype, wedifferently from file-sharing users as well as traditional tele- believe that this study could server as a basis for furtherphone users. understanding and discussing the differences between peer- to-peer file-sharing and peer-to-peer VoIP systems.6. FUTURE WORK We have only scratched the surface of understanding how REFERENCESpeer-to-peer supports VoIP. More generally, interactive ap- [1] BASET, S. A., AND S CHULZRINNE , H. An Analysis of theplications such as peer-to-peer web-caching, VoIP, instant Skype Peer-to-Peer Internet Telephony Protocol. Inmessaging, games etc. may demonstrate different character- Proceedings of the INFOCOM ’06 (Barcelona, Spain, Apr.istics than P2P file-sharing networks and we are interested in 2006).understanding these differences. Measuring existing interac- [2] B HAGWAN , R., S AVAGE , S., AND VOELKER , G. M. Understanding availability. In Proceedings of the IPTPS ’03tive networks including instant messaging networks (AIM, (Berkeley, CA, Feb. 2003).MSN, Yahoo!) and massively multiplayer game networks [3] B ICHLER , G., AND C LARKE , R. V. Preventing Mass(World of Warcraft, Ultima Online) can reveal different user Transit Crime. Criminal Justice Press, Monsey, NY, 1996,behavior. In addition, it would be useful to compare user ch. Eliminating Pay Phone Toll Fraud at the Port Authorityexperience, call setup latency and call quality in Skype and Bus Terminal in Manhattan.other infrastructure-based telephony services including tra- [4] C ALYAM , P., S RIDHARAN , M., M ANDRAWA , W., AND S CHOPIS , P. Performance measurement and analysis ofditional telephone and cellular networks, and VoIP networks h.323 traffic. In Proceedings of the 5th Internationalthat use SIP and H.323 for signaling. Combined, these would Workshop on Passive and Active Network Measurementgive insights about how peer-to-peer networks for such ap- (PAM 2004) (Antibes Juan-les-Pins, France, Apr. 2004).plications should be built and provisioned. [5] C ASTRO , M., C OSTA , M., AND ROWSTRON , A. Debunking some myths about structured and unstructured overlays. In Proceedings of the NSDI ’05 (Boston, MA, May 2005).7. CONCLUSIONS [6] C HAWATHE , Y., R ATNASAMY, S., B RESLAU , L., This paper presents the first measurement study of the L ANHAM , N., AND S HENKER , S. Making gnutellalike p2pSkype VoIP system. From the empirical data we have gath- systems scalable. In Proceedings of SIGCOMM ’03 (Karlsruhe, Germany, Aug. 2003).ered, it is clear that Skype differs significantly from other [7] C HU , J., L ABONTE , K., AND L EVINE , B. N. Availabilitypeer-to-peer file-sharing networks in several respects. Ac- and locality measurements of peer-topeer file systems. Intive clients show diurnal and work-week behavior analogous Proceedings of ITCom ’02 (Boston, MA, July 2002).to web-browsing rather than file-sharing. Stability of the su- [8] C OMBS , G. Ethereal: A Network Protocol Analyzer.pernode population tends to mitigate churn in the network. [9] C ROVELLA , M. E., AND B ESTAVROS , A. Self-similarity inSupernodes typically use little bandwidth even though they world wide web traffic: evidence and possible causes. IEEE/ACM Transactions on Networking 5, 6 (Dec. 1997),relay VoIP and file-transfer traffic in certain cases. 835–846. While the observations above are clear from our data, we [10] F ORD , B., S RISURESH , P., AND K EGEL , D. Peer-to-peermention other observations that are not so clear and are pre- communication across network address translators. In
  6. 6. Proceedings of the 2005 USENIX Annual Technical Conference (Anaheim, CA, Apr. 2005).[11] G ARFINKEL , S. L. VoIP and Skype Security, Jan. 2005.[12] G UHA , S., AND F RANCIS , P. Characterization and measurement of tcp traversal through nats and firewalls. In Proceedings of the 2005 Internet Measurement Conference (New Orleans, LA, Oct. 2005).[13] G UMMADI , K. P., D UNN , R. J., S AROIU , S., G RIBBLE , S. D., L EVY, H. M., AND Z AHORJAN , J. Measurement, modeling, and analysis of a peer-to-peer file-sharing workload. In Proceedings of the SOSP ’03 (Bolton Landing, NY, Oct. 2003).[14] J IANG , W., AND S CHULZRINNE , H. Assessment of voip service availability in the current internet. In Proceedings of the 4th International Workshop on Passive and Active Network Measurement (PAM 2003) (La Jolla, CA, Apr. 2003).[15] L I , J., S TRIBLING , J., M ORRIS , R., K AASHOEK , M. F., AND G IL , T. M. A performance vs. cost framework for evaluating dht design tradeooffs under churn. In Proceedings of the INFOCOM ’05 (Miami, FL, Mar. 2005).[16] L IANG , J., K UMAR , R., AND ROSS , K. W. The kazaa overlay: A measurement study. Computer Networks 49, 6 (Oct. 2005).[17] M ARKOPOULOU , A. P., T OBAGI , F. A., AND K ARAM , M. J. Assessing the quality of voice communications over internet backbones. IEEE/ACM Transactions on Networking 11, 5 (Oct. 2003), 747–760.[18] N EILAN , T., AND B ELSON , K. Ebay to buy internet phone firm for $2.6 billion. The New York Times (Sept. 12, 2005).[19] N OLL , A. M. Cybernetwork technology: Issues and uncertainties. COMMUNICATIONS OF THE ACM 39, 12 (Dec. 1996), 27–31.[20] PAXSON , V., AND F LOYD , S. Wide-area traffic: the failure of poisson modeling. In Proceedings of the SIGCOMM ’94 (London, UK, Aug. 1994).[21] P OUWELSE , J., G ARBACKI , P., E PEMA , D., AND S IPS , H. The bittorrent p2p file-sharing system: Measurements and analysis. In Proceedings of the IPTPS ’05 (Ithaca, NY, Feb. 2005).[22] Q IAO , Y., AND B USTAMANTE , F. Elders know best - handling churn in less structured p2p systems. In Proceedings of the 5th IEEE International Conference on Peer-to-Peer Computing (P2P ’05) (Konstanz, Germany, Aug. 2005).[23] R HEA , S., G EELS , D., ROSCOE , T., AND K UBIATOWICZ , J. Handling churn in a dht. In Proceedings of the USENIX ’04 (Boston, MA, June 2004).[24] ROSENBERG , J., M AHY, R., AND H UITEMA , C. Internet draft: TURN – Traversal Using Relay NAT, Mar. 2006. Work in progress.[25] ROSENBERG , J., W EINBERGER , J., H UITEMA , C., AND M AHY, R. RFC 3489: STUN – Simple Traversal of User Datagram Protocol (UDP) Through Network Address Translators (NATs), Mar. 2003.[26] S AROIU , S., G UMMADI , P. K., AND G RIBBLE , S. D. A measurement study of peer-to-peer file sharing systems. In Proceedings of the Multimedia Computing and Networking 2002 (MMCN’02) (San Jose, CA, Jan. 2002).[27] S EN , S., AND WANG , J. Analyzing peer-to-peer traffic across large networks. In Proceedings of the 2nd ACM SIGCOMM Workshop on Internet measurment (Marseille, France, Nov. 2002).[28] X U , Z., AND H U , Y. Sbarc:a supernode based peer-to-peer file sharing system. In Proceedings of the 8th IEEE Symposium on Computers and Communications (ISCC’03) (Antalya, Turkey, July 2003).[29] Z HAO , B. Y., D UAN , Y., H UANG , L., J OSEPH , A. D., AND K UBIATOWICZ , J. D. Brocade: Landmark routing on overlay networks. In Proceedings of the IPTPS ’02 (Cambridge, MA, Mar. 2002).