See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/2472459
`
`Modeling Player Session Times of On-line Games
`
`Article · June 2003
`
`DOI: 10.1145/963900.963902 · Source: CiteSeer
`
`CITATIONS
`27
`
`2 authors, including:
`
`Wu-chang Feng
`Portland State University
`
`107 PUBLICATIONS   4,160 CITATIONS   
`
`SEE PROFILE
`
`READS
`47
`
`Some of the authors of this publication are also working on these related projects:
`
`Participatory Sensing View project
`
`NetS-NOSS: Towards Dynamically Reconfigurable Mobile, Multimodal Sensing Systems View project
`
`All content following this page was uploaded by Wu-chang Feng on 18 March 2013.
`
`The user has requested enhancement of the downloaded file.
`
`Supercell
`Exhibit 1017
`Page 1
`
`

`

`Modeling Player Session Times of On-line Games
`
`Francis Chang Wu-chang Feng
`OGI@OHSU
`{francis,wuchang}@cse.ogi.edu ∗
`
`ABSTRACT
`One of the most important aspects in determining the global
`traffic characteristics of on-line games is to model the traffic
`behavior of the client. While modeling the client ON-OFF
`times in web traffic has allowed researchers to generate accu-
`rate fractal models of aggregate web traffic, similar work has
`not yet been done for the domain of games. Previous work
`in client traffic behavior has shown that clients consume a
`nearly constant amount of resources when they are on. How-
`ever, the distribution of the length of ON times for clients
`has not yet been established [1]. In this paper, we study the
`player session time distribution over a one-week trace of a
`popular on-line game server. Our results indicate that the
`session ON times of game clients steeply decays over time,
`with a knee at about 15-30 minutes. In addition, we show
`that the session time PDF can be fitted accurately with a
`Weibull distribution with parameters (β = 0.5, η = 20, and
`γ = 0), a result that diverges from previous studies of player
`session times [2].
`
`1.
`
`INTRODUCTION
`With the launches of Microsoft’s Xbox on-line game net-
`work and Sony’s Playstation 2 on-line game network, along
`with the development of emerging massively multi-player
`on-line games that allow for thousands of simultaneous play-
`ers [3], it is clear that gaming traffic will soon grow to scales
`far beyond what is being observed today. While there has
`been a lot of work characterizing aggregate web traffic be-
`havior as well as web client behavior, there has been very
`little work in doing the same for games. Because of the
`∗This work is supported by the National Science Founda-
`tion under Grant EIA-0130344 and the generous donations
`of Intel Corporation. Any opinions, findings, or recommen-
`dations expressed are those of the author(s) and do not nec-
`essarily reflect the views of NSF or Intel.
`
`Permission to make digital or hard copies of all or part of this work for
`personal or classroom use is granted without fee provided that copies are
`not made or distributed for profit or commercial advantage and that copies
`bear this notice and the full citation on the first page. To copy otherwise, to
`republish, to post on servers or to redistribute to lists, requires prior specific
`permission and/or a fee.
`NetGames ’03 May 22-23, 2003, Redwood City, California USA
`Copyright 2003 ACM 1-58113-734-6/03/05 ...$5.00.
`
`Game
`
`Half-life
`Medal of Honor: Allied Assault
`Quake III Arena
`Battlefield 1942
`Unreal Tournament
`Return to Castle Wolfenstein
`Unreal Tournament 2003
`Soldier of Fortune 2: Double Helix
`America’s Army: Operations
`Neverwinter Nights
`
`Players
`
`83121
`7127
`5027
`4473
`4240
`3538
`3052
`2863
`2303
`1818
`
`Table 1: Game popularity Thu Oct 31 12:00 2002
`(Source: Gamespy)
`
`tremendous increase in game traffic [4], it is important to
`understand the traffic demands that this emerging applica-
`tion imparts.
`While not indicative of all on-line games, the class of
`games known as “first-person shooters” has clearly domi-
`nated much of the observed gaming traffic. In this paper,
`we analyze the session behavior of game clients playing on
`a popular FPS game server and develop a source model for
`their ON times. We note that previous work has shown that
`when such clients are active, they impart almost a constant
`bit-rate load on the network.
`Section 2 describes a detailed game server trace we per-
`formed in order to analyze the session behavior of players in
`on-line games. Section 3 analyzes the session times of play-
`ers in the trace and develops a source model that describes
`session times of players in the trace. Section 4 concludes
`with a discussion of the implications that these results have
`on traffic characterization of on-line games.
`
`2. BACKGROUND
`In order to study a representative on-line game, we chose
`to characterize session behavior of a busy Counter-Strike
`server. Currently, Counter-Strike (a Half-Life modification)
`is the clearly the most dominant on-line game with the
`largest service footprint of any game at 35,000 servers and
`over 4.5 billion player minutes per month [5]. Table 1 shows
`a snapshot of the top on-line multi-player action games taken
`
`Supercell
`Exhibit 1017
`Page 2
`
`

`

`Half-Life mod
`
`Players
`
`Counter-Strike
`Day of Defeat
`Team Fortress Classic
`Deathmatch
`The Specialists
`Firearms
`SvenCo-op
`Vampire Slayer
`Front Line Force
`Action Half-Life
`
`74877
`3303
`3010
`557
`321
`302
`87
`83
`64
`55
`
`Table 2: Half-Life mod popularity Thu Oct 31 12:00
`2002 (Source: Gamespy)
`
`Start Time
`Stop Time
`Total Time
`
`Maps Played
`Player sessions
`
`Apr 11 08:55:04 2002
`Apr 18 14:56:21 2002
`7 d, 6 h, 1 m, 17.03 s
`(626,477 sec)
`339
`16030
`
`Table 3: mshmro trace information
`
`in late October 2002. The table was generated via the
`GameSpy Arcade application [6], an on-line service used
`by players to find suitable game servers. GameSpy Ar-
`cade tracks and reports the number of players playing in-
`dividual on-line games in real-time. As the table shows,
`Half-Life and its variants account for an order of magnitude
`more players than the next nearest game. The dominance
`of Half-Life goes back as far as 2000 [4], where measure-
`ments indicated that the application was generating a large
`percentage of all observed UDP traffic behind DNS and Re-
`alAudio traffic. Table 2 shows that of the people playing
`Half-Life variants, an overwhelming majority of them are
`playing Counter-Strike.
`Counter-Strike is architected as a client-server applica-
`tion with multiple clients communicating and coordinating
`with a central server that keeps track of the global state of
`the game. In the game, two teams continuously play back-
`to-back rounds of several minutes in duration. Each team
`attempts to complete their objectives and foil those of the
`other team during the round. The round ends when one
`team manages to complete their mission, when one team
`eliminates the other team entirely, or when time runs out.
`When players are eliminated from a round, they become
`spectators until the next round. During this time, the player
`can shadow another player that has not been eliminated.
`The game itself is played on a variety of maps which rotate
`based on how the server is configured. Typically, maps are
`rotated every 30 minutes, allowing for over 10 rounds to be
`played per map. Depending on the hardware and configu-
`ration, a Counter-Strike server can support up to 32 simul-
`taneous players. Previous work has shown that the network
`
`load induced by clients when they are active is constant and
`is a result of the saturation of the narrowest last-mile link [1].
`Given that network load for a client is relatively constant
`when connected, it is interesting to find out what the session
`times for players is in order to more accurately construct a
`source model for game traffic. To characterize session times
`for Counter-Strike, we used a trace from a popular Counter-
`Strike (version 1.3) server located at cs.mshmro.com [7, 1].
`It is important to note that the session time distributions
`are heavily dependent on the quality of the game server. Be-
`cause of this, the server was set up to provide an experience
`that was as positive as possible. In particular, at the time
`of the trace, cs.mshmro.com had a high-speed connection to
`the Internet (OC-3 to Internet2, dual T3s to non-Internet2
`sites) and a fast machine to run on (Dell Dimension 4300,
`Pentium 4, 1.8GHz, 512MB). In addition, maps were rotated
`every 30 minutes to prevent player boredom and modules for
`eliminating cheating and team killing [8, 9, 10, 11, 12] were
`added. While finding a server configuration that maximizes
`session times is not an exact science, our own personal expe-
`riences playing on the server indicated a configuration that
`encouraged long-term play. This was evidenced in the fact
`that the server quickly became heavily utilized 24/7 with
`connections arriving from all parts of the world irrespective
`of the time of day [4, 2, 13].
`Table 3 summarizes the key characteristics of the trace.
`The trace covered over a week of continuous operation. Over
`300 maps were played during this time frame and more than
`16000 user sessions were established. Due to the popularity
`of the server, each user averaged almost 3 sessions for the
`week and more than 8000 connections were refused due to
`the lack of open slots on the server. Note that for our study,
`these refused connections were not counted as legitimate
`user sessions.
`
`3. EVALUATION
`Given the trace above, we extracted the total session time
`of each player session contained in the trace. Figure 1(a)
`plots the session time histogram of the trace in unit incre-
`ments of a minute. The figure shows, quite suprisingly, that
`a significant number of players play only for a short time be-
`fore disconnecting and that the number of players that play
`for longer periods of time drops sharply as time increases.
`Note that in contrast to heavy-tailed distributions reported
`for most source models for Internet traffic, the session ON
`times for game players is decidedly not heavy-tailed. To fur-
`ther illustrate this, Figure 1(b) shows the cumulative density
`function for the session times of the one-week trace. As the
`figure shows, the probability of having a particular session
`time drops sharply with more than 99% of all sessions lasting
`less than 2 hours.
`In order to develop accurate source models for game traf-
`fic, it is useful to match a distribution to the probability
`density function of the session times. As the session times
`model “player” lifetimes, the PDF can be closely matched
`to a Weibull distribution, one of the most common distri-
`butions used to model lifetime distributions in reliability
`
`Supercell
`Exhibit 1017
`Page 3
`
`

`

`(a) Histogram
`
`(b) CDF
`
`Figure 1: Session time results for mshmro trace
`
`Figure 2: Fitted Weibull distribution on session time
`PDF of mshmro trace
`
`Figure 3: Player failure rates for individual session
`times
`
`engineering [14]. It is not surprising that this distribution
`fits, since quitting the game can be viewed as an attention
`“failure” on the part of the player. The generalized Weibull
`distribution has three parameters β, η, and γ and is shown
`below.
`
`−( T −γ
`η ( T−γη )β−1e
`
`η )β
`f (T ) = β
`In this form, β is a shape parameter or slope of the distri-
`bution, η is a scale parameter, and γ is a location parameter.
`As the location of the distribution is at the origin, γ is set
`to zero, giving us the two-parameter form for the Weibull
`PDF.
`
`
`f (T ) = β
`
`−( T
`η ( Tη )β−1e
`η )β
`Using a probability plotting method [14], we estimated the
`shape (β) and scale (η) parameters of the session time PDF.
`
`As Figure 2 shows, a Weibull distribution with β = 0.5,
`η = 20, and γ = 0 closely fits the PDF of measured session
`times for the trace.
`Note that this result is in contrast to previous studies that
`have fitted an exponential distribution to session-times of
`multiplayer games [2]. Unlike the Weibull distribution which
`has independent scale and shape parameters, the shape of
`the exponential distribution is completely determined by λ,
`the failure rate. Due to the memory-less property of the ex-
`ponential distribution, this rate is assumed to be constant.
`Figure 3 shows the failure rate for individual session dura-
`tions over the trace. As the figure shows, the failure rate is
`not constant for shorter session times, thus making it diffi-
`cult to accurately fit it to an exponential distribution. In-
`explicably, the failure rate is higher for flows of shorter du-
`ration. While it is difficult to pinpoint the exact reason for
`
`1600
`
`1400
`
`1200
`
`1000
`
`800
`
`600
`
`400
`
`200
`
`Number of players
`
`0
`
`0
`
`15
`
`30
`
`45
`
`60
`Minutes
`
`75
`
`90
`
`105
`
`120
`
`1
`
`0.9
`
`0.8
`
`0.7
`
`0.6
`
`0.5
`
`0.4
`
`0.3
`
`0.2
`
`0.1
`
`Probability
`
`0
`
`0
`
`30
`
`60
`
`90
`
`120
`Minutes
`
`150
`
`180
`
`210
`
`240
`
`PDF of mshmro player session times
`Weibull b=0.5, h=20, g=0
`
`0.15
`
`0.1
`
`0.05
`
`Probability
`
`0
`
`0
`
`15
`
`30
`
`45
`
`60
`Minutes
`
`75
`
`90
`
`105
`
`120
`
`0.2
`
`0.15
`
`0.1
`
`0.05
`
`Player failure rate
`
`0
`
`0
`
`20
`
`40
`
`60
`
`Minutes
`
`80
`
`100
`
`Supercell
`Exhibit 1017
`Page 4
`
`

`

`releaseview.cfm?ReleaseID=1050, 2003.
`[6] GameSpy.com, “GameSpy: Gaming’s Home Page,”
`http://www.gamespy.com/, 2002.
`[7] mshmro.com, ,” http://www.mshmro.com/.
`[8] Half-Life Admin Mod Developers, “Half-Life Admin
`Mod Home,” http://www.adminmod.org/.
`[9] Cheating-Death Developers, “Cheating-Death Home,”
`http://www.cheating-death.com/.
`[10] CSCop Developers, “Project Info - CSCop for
`Counter-Strike,”
`http://sourceforge.net/projects/cscop.
`[11] CSGuard Developers, “CSGuard Home,”
`http://www.olo.counter-strike.pl/.
`[12] UDPSoft.com, “HLDS Ping Booster,”
`http://www.udpsoft.com/.
`[13] T. Henderson, “Latency and User Behaviour on a
`Multiplayer Game Server,” in Networked Group
`Communication, 2001, pp. 1–13.
`[14] ReliaSoft Corporation, “Life Data Analysis and
`Reliability Engineering Theory and Principles
`Reference from ReliaSoft,”
`http://www.weibull.com/lifedatawebcontents.htm,
`2003.
`[15] AMX Mod Developers, “AMX Mod Server Plugin,”
`http://amxmod.net/.
`
`
`View publication statsView publication stats
`this, it could be attributed to the fact that Counter-Strike
`servers are notoriously heterogeneous. Counter-Strike hap-
`pens to be one of the most heavily modified on-line games
`with support for a myriad of add-on features [8, 15]. Short
`flows could correspond to players browsing the server’s fea-
`tures, a characteristic not predominantly found in other
`games. In addition, this particular server was a friendly-fire
`server, a realistic setting in which team members are able to
`shoot each other. To protect the game from rampant team-
`killing, a plugin was used to kick players for continuously
`killing their teammates. It may be possible that new play-
`ers not cognizant of this plugin, were quickly kicked from
`the server. As part of future work, we hope to characterize
`session duration distributions across a larger cross-section of
`games in order to see how distributions vary between games
`and game genres.
`
`4. CONCLUSION
`Unlike many source models used to predict Internet traf-
`fic, player session “ON” times follow a distribution that is
`not heavy-tailed. Session times, in fact, follow a simple
`Weibull distribution, a distribution that is commonly used
`to model lifetime distributions for reliability analysis. This
`work, along with previous studies showing a constant re-
`source demand for active clients [1], allows for the construc-
`tion of a partial source model for FPS-based on-line games.
`In order to more fully develop this source model, we are
`currently examining the inter-session arrival times of par-
`ticular players. While previous work [2] and our own traces
`can elucidate inter-session arrivals for players to a particular
`server, such an analysis is inherently inaccurate since players
`often play on many different servers. We are attempting to
`extract such information via global player location services
`such as those provided by GameSpy [6] and by attempting
`to gain access to traces of centralized game authentication
`servers to develop a more accurate model for player “OFF”
`times. Such a model will allow for the construction of a
`complete source model for FPS games.
`
`5. REFERENCES
`[1] W. Feng, F. Chang, W. Feng, and J. Walpole,
`“Provisioning On-line Games: A Traffic Analysis of a
`Busy Counter-Strike Server,” in Proc. of the Internet
`Measurement Workshop, November 2002.
`[2] T. Henderson and S. Bhatti, “Modelling User
`Behavior in Networked Games,” in ACM Multimedia,
`2001, pp. 212–220.
`[3] GameSpy.com, “What’s This World Coming To? The
`Future of Massively Multiplayer Games,”
`http://www.gamespy.com/gdc2002/mmog.
`[4] S. McCreary and k. claffy, “Trends in Wide Area IP
`Traffic Patterns: A View from Ames Internet
`Exchange,” in Proceedings of 13th ITC Specialist
`Seminar on Measurement and Modeling of IP Traffic,
`September 2000, pp. 1–11.
`[5] AMDZone, “Valve Releases Hammer Port of
`Counter-Strike Server,” http://www.amdzone.com/
`
`Supercell
`Exhibit 1017
`Page 5
`
`