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

1. An Experimental Study of the Skype Peer-to-Peer VoIP System
Saikat Guha Neil Daswani, Ravi Jain
Cornell University Google
saikat@cs.cornell.edu {daswani, ravijain}@google.com
ABSTRACT sharing usage? Gummadi et al. [13], find that file-sharing
Despite its popularity, relatively little is known about the traffic users are patient when their file-searches succeed and leave
characteristics 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. In
VoIP traffic conducted over a five month period, where over 82 mil- contrast, we find that Skype users regularly run the client
lion 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 consider
This 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 to
i.e., without knowing the internals or specifics of the Skype system provide a global VoIP service, or can a global VoIP service be
or 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 peers
Skype 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, in
number of active clients shows diurnal and work-week behavior, §2, we shed light on some design choices in the proprietary
correlating 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 node
stable; thus node churn, a significant concern in other systems, dynamics and churn in Skype’s peer-to-peer overlay, and the
seems less problematic in Skype. The typical bandwidth load on a
supernode 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 of
that may be useful for future design and modeling of such sys- the present paper. Altogether, we find evidence that Skype
tems. These results also imply that the nature of a VoIP P2P system is fundamentally different from the peer-to-peer networks
like 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 to
1. 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 Skype
fast supplementing email in both enterprise and home net- user (if the recipient agrees)1 . Skype also offers paid services
works. Skype is an application that provides these VoIP and that allow Skype users to initiate and receive calls via regular
IM 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 overlap
the Skype P2P VoIP network and obtain significant amounts of technical staff, and that much of the technology in Skype
of data. This data was collected from September 1, 2005 to was originally developed for KaZaA. Network packet level
January 14, 2006 at Cornell University. While measurement analysis of KaZaA [16] and of Skype [1] support this claim
studies 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 research
networks and from traffic in traditional voice-communication in [29, 28, 6, 5]. Typically, supernodes maintain an overlay
networks. For example, how does the interactive-nature of
VoIP 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. 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 supernodes
amongst them based on some criteria. Ordinary nodes issue between Jul. 25, 2005 and Oct. 12, 2005. Each client caches
queries 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 the
ment to examine the basic operation and design of the Skype addresses in the cache to a list. Our script then replaces
network in some more detail. We ran two Skype clients the cache with a single supernode address from the list such
(version 1.1.0.13 for Linux) on separate hosts, and observed that the client is forced to pick that supernode the next time
the destination and source IP addresses for packets sent and the client is run. The script starts the client and waits for
received in response to various application-level tasks. We it to download a fresh set of supernode addresses from the
observed that in Skype, ordinary nodes send control traffic supernode to which it connects. The script then kills the
including availability information, instant messages, and re- client causing it to flush its supernode-cache. The cache is
quests for VoIP and file-transfer sessions over the supernode processed again and the entire process repeated; the result is
peer-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 supernode
between 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 Skype
clients behind different NATs. We observed that if one client users each time the client is run, as reported by the Skype
is 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 given
session. If both clients are behind NATs, Skype uses STUN- time. We wrote a tool that sends application-level pings to
like NAT traversal [25, 10] to establish the direct connection. supernodes; the tool replays the first packet sent by a Skype
In the event that the direct connection fails, Skype falls back client to a supernode in its cache, and waits for an expected
to a TURN-like [24] approach where the media session is response. For each snapshot, we perform parallel pings to
relayed by a publicly reachable supernode. This latter ap- a fixed set of 6000 nodes randomly selected from the set of
proach 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 at
that Skype employs to service VoIP and file transfer requests 30 minute intervals for one month beginning Sep. 12, 2005.
is 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/UDP
nodes are promoted to supernodes. In an experiment we con- header fields including the source and destination IP address
ducted, we ran several Skype nodes in various environments and port, and packet lengths.
and waited two weeks for them to become supernodes. A
Skype node behind a saturated network uplink, and one be- 4. CHARACTERIZING SKYPE’S NETWORK
hind a NAT, did not become supernodes, while a fresh install Churn in P2P networks, the continuous process of nodes
on a public host with a 10 Mbps connection to the Internet joining and leaving the system, increases routing latency as
joined the supernode network within minutes. Consequently,
some overlay traffic is routed through failed nodes, while
it 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 maintenance
system. We did not test additional criteria such as a history
traffic. Churn has been studied extensively in peer-to-peer
of long session times, or low processing load as suggested file-sharing networks [26, 27, 2, 13, 7]; the consensus is that
in [28]. As we show later, the population of supernodes se-
churn can be high. The session time of a node is defined as
lected by Skype, apparently based on reachability and spare
the time between when a node joins the network, and then
bandwidth, 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 few
heterogeneity is leveraged to control churn, not just cope with
minutes, and that frequent updates are needed to maintain
it. consistency [23].
In this section, we study churn in Skype’s supernode P2P
3. 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, and
observed the network activity of a Skype supernode for 135 that supernodes demonstrate diurnal behavior causing median
days (Sep. 1, 2005 to Jan. 14, 2006). We ran version 1.2.0.11 session times of several hours. Further, we find that session
of 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 supernodes
received 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 left
2 We overload NAT to mean NATs and firewalls in the rest of the tracks daily variations in client and supernode populations
paper. from Sep. 18 to Oct. 4, while the plot on the right zooms-in on

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 contributes
Figure 1: (a) Percentage of all nodes, and supernodes active at any 15–25% of supernodes, with a peak contribution around noon
time. (b) Geographic distribution of supernodes as observed over CST. Similarly, Asia contributes 20–25% and peaks around
the duration of our trace. its mid-day. Combined with the lower weekend-usage from
the previous graph, there is evidence to conclude that Skype
hourly variations on Sep. 22. As mentioned in Section 3, the usage, at least for those nodes that become supernodes, is
number of online users is reported by the official Skype client, correlated with normal working hours. This is different
while online supernodes are determined through periodic from P2P file-sharing networks where users download files in
application-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 working
usage 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 in
ations 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 leaving
parison, 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 same
and between 30–40% of them are online at any given time. experiment is 5.5 hours, as shown in Figure 3. The median
While client population varies by over 40% on any given is higher than reported in previous studies of file-sharing
day, supernode population is more stable and varies by under networks [26, 27, 2, 13, 7]; however, these studies measure
25%. 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 in
working 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. mentary CDF is not a strict straight line, i.e., a strict power 1
law, it may be approximated as such for our data. 0.9
The nature of this distribution suggests that both arrivals 0.8
and departures in Skype cannot be modelled as a Poisson
or uniform processes. Consequently, results from past work 0.7
that models churn as fixed-rate Poisson processes [23, 5, 15] 0.6
may be misleading if directly applied to Skype.
CDF
0.5
One way to model churn in Skype is to model node arrival
0.4
as a Poisson process with varying hourly rates (higher in
the morning), and picking session times from a heavy-tailed 0.3
distribution, similar to the approach used in in [22]. Never- 0.2
theless, since there is little churn in the first place (more than 0.1 Control and IM traffic
95% of supernodes persist from one thirty-minute snapshot All traffic
Relayed VoIP/Filetransfer
to the next), we expect periodic updates with an update-rate 0
30 100 300 1k 3k 10k 30k 100k
chosen accordingly to perform well [23]. As an optimiza- Bandwidth (bps)
tion, reactive updates can be used to conserve bandwidth at
night, 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 sent
dreds 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 per
companies and wireless carriers who rely on expensive, ded- second and an overall one-way bytes-sent ratio greater than
icated, 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 the
text of VoIP. We find that Skype supernodes incur a small time. This value is smaller than we expected; it is explained
network 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. For
Skype 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-sharing
less 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 processing
teristics. 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 insights
ode 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 only
resents a convenient looking-glass into a global VoIP/IM for relayed VoIP/file-transfer sessions and not IM or direct
network. However, this study is limited by the Skype’s en- sessions. This potentially introduces an unavoidable bias
cryption of control and data messages. The data we have in our user population. Second, this causes us to miss an
used for studying VoIP in Skype is extracted from the dataset estimated 85% of the data (based on [12]), resulting in only
obtained by Expt. 3-5 described previously. However, we 1121 data points for 135 days. Nonetheless, we believe
are only able to examine sessions that involve supernodes, that even these preliminary results show interesting trends
and also must make a heuristic estimate of which sessions and, to the best of our knowledge, represent the first publicly
are 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 user
sharing 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 file-sharing.
is required to validate these observations. In particular, it VoIP session length behavior. Figure 5(b) shows that the
is possible that calls that are relayed by supernodes have median Skype call lasted 2m 50s, while the average was
different 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. 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 VoIP
conversation 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 that
acquired dialing codes to make long-distance calls for free, indicate that it is possible that Skype calls are significantly
would place calls that lasted 9 minutes on average, while longer than calls in traditional telephone networks, while
legitimate 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 to
size is similar to documents, presentations and photos, and validate these preliminary observations.
is much smaller than audio or video files in file-sharing net- Overall, we present measurement data useful for designing
works [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, we
differently from file-sharing users as well as traditional tele- believe that this study could server as a basis for further
phone 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 REFERENCES
peer-to-peer supports VoIP. More generally, interactive ap- [1] BASET, S. A., AND S CHULZRINNE , H. An Analysis of the
plications such as peer-to-peer web-caching, VoIP, instant Skype Peer-to-Peer Internet Telephony Protocol. In
messaging, 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 ’03
tive 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 Authority
experience, 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 of
ditional telephone and cellular networks, and VoIP networks h.323 traffic. In Proceedings of the 5th International
that use SIP and H.323 for signaling. Combined, these would Workshop on Passive and Active Network Measurement
give 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 p2p
Skype 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. Availability
peer-to-peer file-sharing networks in several respects. Ac- and locality measurements of peer-topeer file systems. In
tive 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 in
Supernodes 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-peer
mention other observations that are not so clear and are pre- communication across network address translators. In

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).