Network Analysis of Counter-strike and Starcraft
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Network Analysis of Counter-strike and Starcraft
Mark Claypool, David LaPoint, Josh Winslow
{ claypoo1)Qcs. wpi .edu
Computer Science Department
Worcester Polytechnic Instit u te
100 Institute Road, Worcester, MA 01609, USA
Abstract net. While the existing empirical studies have been
Network games are becoming increasingly popular, valuable in helping understand Internet performance,
but their trafic patterns have received little attention they are not sufficient for characterizing network game
from the academic research community. A better un- traffic since network games have different application
derstanding of network game t r a f i c can lead t o more requirements, and hence different network traffic char-
effective network architectures and more realistic net- acteristics, than the majority of Internet traffic.
work simulations. W e gathered t r a f i c traces from live Unlike typical Internet traffic, network games typi-
Internet games of Counter-strike and Starcraft, repre- cally do not use reliable Internet transport protocols.
sentative games from two different gaming genres. In The majority of reliable Internet traffic runs on top of
this paper, we analyzed the network trafic patterns and TCP, which guarantees sequential, in order arrival of
bandwidth consumption t o better understand how such all data by retransmitting lost packets. However, the
games impact the network. B o t h games were found additional time required for retransmissions and the
t o have very low bandwidth usage and packets signifi- window-based data rate imposed by TCP can be un-
cantly smaller than typical Internet applications. Star- acceptable to network games since interactive games
craft games showed consistent trafic behavior across often have strict delay constraints [2]. Instead of TCP,
games while the trafic patterns f o r Counter-strike var- network games often use UDP since UDP packets are
ied widely, especially f o r the server. In particular, not guaranteed t o reach their destinations, and thus do
Counter-strike clients and servers often send bursts of not require a mechanism for retransmissions. Unfortu-
very small packets. Since current Internet routers are nately, UDP does not provide any built-in congestion
typically designed f o r large transfers with large packets, control, presenting the risk of congestion collapse as
there m a y be opportunities to improve network archi- the fraction of unresponsive UDP traffic increases [43.
tectures t o better manage and support game trafic. Despite the very real burden that network games
may impose on network bandwidth and the massive
1 Introduction growth and large user base, issues related to the ef-
The rapid growth in the connectivity of the Internet fects of these games on network congestion have been
and the fall in computer hardware prices has fueled the largely neglected in both academic and industry publi-
growth of multi-player gaming over the Internet. In cations. Industry articles are more concerned with the
addition to computers, the latest generation of con- management aspects of game development or focus on
soles systems from Nintendo, Sony, and Sega all in- latency and maximum bandwidth issues only [3,6,11],
clude network support for multi-player game play on whereas attention to detail like minimizing network
the Internet. The increase in Internet gaming is also a load suffer due to the lack of economic pressure. Net-
result of change in developer focus from single player work gaming has not been a traditional research field
games to multi-player games. in academia as network games are viewed primarily
Over the years, there have been a number of studies as a diversion for students, rather than a practical
measuring the performance of Internet backbones and problem for computer scientists. Among other things,
end-hosts [13, 141, as well as detailed studies on the a better understanding of network game traffic can
performance of Web clients [5, 81 and streaming me- help design networks and architectures that more ef-
dia [7, 10, 151. However, to the best of our knowledge fectively accommodate their traffic footprints. Fur-
there have not been wide-scale empirical measurement thermore, careful empirical measurements of network
of multi-player network game traffic across the Inter- games can provide the data required for accurate simu-
0-7803-7893-8/03/$17.00 0 2003 IEEE 263lations, a typical tool for evaluating network research. User Dungeons, first pioneered in the 70s6. MMORP
The goal of this research is to analyze the network games provide a mechanism for character advance-
traffic of two of the most popular games on the Inter- ment, large lands to travel across, and other players
net. We chose two representative games from the most with whom to interact. The three biggest MMORP
popular genres, First Person Shooters and Real Time games, Asheron's Call7, Ultima Online', and Ev-
Strategies (see Section 2.1), with the intent of repre- erquestg, claim to have nearly 1 million subscribers
senting a larger range of data than a single genre would combined.
likely provide. We used a network sniffer on both game RTS games, first made popular by Dune 21°, are
servers and game clients participating in live Internet generally characterized by resource collection, unit
games in order to capture realistic network traffic. We construction, and battles that consist of large num-
then analyze the distribution of packet sizes, packet in- bers of animated soldiers going through a repetitive,
terarrival times, differences between client and server animated attack. All of these actions happen in real-
bandwidths and impact of number of players on net- time, unlike earlier strategy games such as Civiliza-
work traffic. tion", in which the player could take as much time
The rest of this paper is laid out as follows: Sec- as needed to plan the next turn. Since Dune 2, there
tion 2 details our approach in selecting network games have been several more RTS games released, one of the
and running our experiments; Section 3 provides re- more popular being Starcraft12. Currently, the typi-
sults and analysis of the data traces collected during cal number RTS fans playing Starcraft on an average
the experiments; Section 4 briefly describes how re- night numbers about 20,000 players.
sults from Section 3 could be used to simulate network We selected Starcraft, a RTS game, and Counter
games; and Section 5 summarizes our conclusions and Strike, a FPS game, because they were familiar, best
presents possible future work. selling games in different genres. While we wanted
to study Asheron's Call, a MMORP game, the time
2 Approach it took to setup and run our experiments precluded
We employed the following methodology: select studying another genre. We leave that as future work.
network games for study (Section 2.1), run the selected
games in an instrumented environment (Section 2.2),
2.1.1 The Starcraft G a m e Environment
and analyze network traffic the games generate (Sec-
tion 3). Starcraft13 is a RTS game that has players construct
2.1 Selecting Games buildings and fighting units, and issue commands that
Broadly, the most popular categories of red-time cause the units to move, engage enemy units, and sim-
network games are First Person Shooters (FPS) ilar tasks. Games are played on one of many possible
and Massively Multi-player Online Role Playing maps, either provided with the game or custom built
(MMORP), followed closely by Real Time Strategy by users. There are three races from which a player
(RTS). can choose, and each has a balanced set of advan-
FPS games, first made popular by Doom1, have tages and disadvantages over the others. There are
the player view the world through the perspective a number of ways in which players can be competi-
of a character (the first person) and have the player tively grouped. In a free-for-all game, all pIayers vie
to have the last remaining army on the map. Players
move around slaying monsters and other players, with
an amalgamation of ranged weaponry (the shooter). can also team up against each other and/or against
On an average night, there are well over 10,000 artificially intelligent computer-controlled players in
myriad ways.
servers Half-Life2 games supporting over 40,000 play-
~ . FPS games support slightly smaller user
e r ~ Other For our experiments, all games were structured so
populations4. that there were two teams of equivalent sizes: 1 vs.
MMORP games, first made popular by Ultima On- http://www.legendmud.org/raph/gaming/book.html
line's release in 19965, are similar to graphical Multi- 'http:/ /www.zone.com/asheronscall/, averages 12,000 play-
ers per night
http://www.idsoftware.com/games/doom/doom-final/ *http://www.uo.com, over 300,000 total subscribers
2http://half-life.sierra.com/ ghttp://everquest.station.sony.com/,over 410,000 total sub-
3Average of 3 randomly selected nights between 10/2/01 and scribers
12/15/01. lohttp://www.dune2k.com/duniverse/dune2/
4Tribes 2, based on an average over 3 randomly selected "http://www.civ3.c0m, the second sequel of Civilization
nights between 10/16/01 and 12/15/01. 12http://www.blizzard.com/worlds-starcraftshtml
http://www.uo.com '3http://www.blizzard.com/worlds-starcraft.shtml
2641, 2 vs. 2, 3 vs. 3, 4 vs. 4. The local player, ana- maps used were de-dust, de-aztec, and cs-assault,
lyzed in detail in Section 3, played as the same race in all from the standard Counter-strike install. The
each game, and employed the same building strategy Counter-strike server we connected to was located on
throughout. The games were played using Starcraft: the WPI network.15
Brood War, version 1.7. The local player logged on to 2.2 Running Games
b a t t l e . net using the USEAST gateway, and created We used Commview (version 2.6, build 103)" to
the game sessions. Each game type was top players capture all network traffic related to the games ran.
vs. bottom players. The same map, called Big Game Commview is a robust network packet sniffer with the
Hunters found in the maps/broodwar/webmaps direc- ability to filter packets, compute statistics; and gener-
tory from where the game was installed, was used for ate reports periodically.
each game. . All Starcraft data was collected on the machine:
2.1.2 The Counter-Strike Game Environment Intel Pentium I11 80Omhz processor with lOOmhz FSB
512 megabytes PC-100 SDRAM
nVidia geForce2 3d v/64 Mbytes of DDR SDRAM
Counter-strike14 is a modification t o the FPS Half- Ultrawide SCSI hard drive interface
Life that is distributed free over the Internet for own- lObaseT nic connected to 608/108 kbps DSL modem
ers of Half-Life, or as a retail product in most game Windows 98B Operating System
stores. Counter-strike puts the players in the role of
either a terrorist attempting to hold hostages, blow up All Counter-strike data was collected on the ma-
landmarks, or assassinate a VIP or a counter-terrorist chine:
agent trying to thwart the terrorists. To play, each
AMD Athlon 8OOmhz processor with 200mhz FSB
player must connect to a server, usually located on an
256 megabytes PC-100 SDRAM
end-host on the Internet. When one or more other nVidia geforce 3d u/32 Mbytes of DDR S D M
players join the same server, a game begins. ATA-66 hard drive interface
All Counter-strike games are played on a pre-loaded lObaseT nic connected to WPI LAN
Windows 98 v4.10.98 Operating System
map, with each map having its own set of victory ob-
jectives. Most objectives have the counter-terrorists
attempting to rescue a set of hostages from close to 3 Results and Analysis
where the terrorists start, or the terrorists attempting This section presents results and analysis for band-
to plant a bomb close to where the counter-terrorists width (Section 3.1) and packets (Section 3.2), for both
start. Typically, each map is played several times, Starcraft and Counter-strike.
with each time being called a round, lasting several 3.1 Bandwidth Analysis
minutes. Each round ends either when the victory ob- Figure 1 depicts the bandwidth sent by each player
jectives are met, time runs out, or when one team has for 2 player, 4 player, 6 player and 8 player Star-
been totally eliminated. At the start of each round, craft games. The amount of bandwidth sent is lin-
both sides are allowed to buy weapons and ammuni- ear with the number of game players with a 2-player
tion with the money they earned from the previous game sending around 650 bytes/second and each pair
rounds. The better a team did in the previous round, of additional players adding about 1500 bytes/second
the more money they have t o spend. Once each team of data. Also, as the number of players increases, so
is equipped, they attempt to completely wipe out the does the variance in bandwidth sent (standard error
other team with their weaponry or complete the ob- of 32% for a 2-player game and 41% for an &player
jective, although the former outcome ends far more game).
rounds than the latter. Figure 2 depicts the bandwidth sent by a Counter-
Controlling the environment in a real Counter- strike game client using the cs-assault map for the
strike game is difficult because players can join and first 20 minutes and then switching to the de-aztec
leave as they please. We chose to analyze several map. The average bandwidth sent is around 2800
different maps to see the effect the campaign has on bytes/second. There is considerable variation in the
network traffic. The number of players was not con- bandwidth per second (standard error of l l O % ) , with
trolled, as players joined and left the server in a typi- a noticeable drop in bandwidth at 1100 seconds when
cal Counter-Strike gaming fashion. We used Counter- there was a map change.
strike version 1.3 on Half-life version 1.1.0.8. The ~-
l5 http://www.wpi.edu/Admin/Netops/MRTG/
''http://www.counter-strike.net 16http://www.tamos.com/products/commview/
265effect of game size on outgoing bandwidth
goo0 I I I I I
I
""V"
I I I I I
I I I I I
7flilIl
I-
J . I I = I* I I
6000
c
0
5s mm n
*2 players
4 players
e
U
* 6 players
- 8 players
1OM
- .
0 100 200 300
lime (seconds)
Figure 1: Starcraft Bandwidth Sent for Different Game Sizes
cs-assault to de-aztec client bandwidth
"1
Figure 2: Counter-strike Bandwidth Sent by Client
266Figure 3 depicts the bandwidth sent by the server 300 milliseconds, as evidenced by the gradual slope.
for the same Counter-strike 3-person game. The av- Counter-strike, on the other hand, has about 30% of
erage bandwidth is around 6470 bytes/second, with its packets arrive in large bursts.
considerable variation in the bandwidth per second Overall, both Counter-strike and Starcraft send
(standard error of 135%). The map change occurs at considerably smaller packets than the typical Internet
about 1110 seconds, depicted by the drop in band- traffic packet size of over 400 bytes [9]. The num-
width. ber of players does not have a noticeable effect on
Overall, Starcraft shows considerably more band- the packet sizes for either Starcraft or Counter-strike.
width consistency than either the Counter-strike client Starcraft packet sizes are consistent throughout the
or the Counter-strike server. The largest Starcraft game, but Counter-strike packet sizes, on the other
game possible, an 8-player game, is still sufficiently hand, are greatly influenced by the action in the game.
optimized to play at modem speeds. Counter-strike Counter-strike sends periodic bursts of small packets
clients can similarly play at modem speeds (overall, while Starcraft send packets at a much steadier rate.
our clients averaged about 1200 bytes/second), but Since current Internet routers are designed for large
since Counter-strike servers can support a much higher transfers with large packets, there may be opportuni-
number of players than Starcraft (games with 20 play- ties to improve network architectures to better manage
ers are not uncommon), server bandwidth may become and support game traffic.
constrained over a modem.
4 Simulation of Network Game Traffic
3.2 Packet Analysis Empirical experiments with network games are of-
Figure 4 depicts a cumulative density function ten difficult because of variable network conditions
(CDF) of packet sizes (both inbound and outbound) and the costs involved with deploying large num-
for Starcraft and Counter-strike. Both games pro- bers of network game clients under controlled con-
duce small packets, with the median packet size for ditions. However, by using simulators, such as NS-
Starcraft about 120 bytes and the median packet for 2 [12], the previous section results may be useful for
Counter-strike clients about 160 bytes. Both Starcraft network game protocol designs, new network router
and the Counter-strike client exhibit a naxrow range queue management disciplines that react to network
of small packets. The Counter-strike server, however, game traffic, and understanding interactions between
has a much heavier tail, with 10% of all packets being network game traffic and traditional traffic. We briefly
over 500 bytes. sketch out the design of traffic generators that could
While a Starcraft game has very little deviation in be used in such simulations.
terms of packet size or bandwidth used throughout First, separate traffic generators should be chosen
the run of one game, Counter-strike, especially from for Starcraft and Counter-strike, with Counter-strike
the viewpoint of the server, has a non-uniform, but traffic generators separated into client and server com-
distinct network pattern. Figure 5 depicts the packet ponents. One can model an overall network game
sizes sent over time for a 3-player Counter-strike game client by selecting an RTT based on average condi-
using the cs-assault map for the first 23 minutes and tions seen in [2]. Packet interarrivals can be proba-
then switching to the de-aztec map. Rounds can be bilistically selected based on data in Figure 6 and the
distinguished by the slow decline, caused by players packet sizes can be selected based on data in Figure 4.
dying, in the packet sizes sent from the server. For Once developed, bandwidth consumption can be vali-
example, one round goes from approximately 500s to dated using the graphs in Figures 1, 2 and 3.
650s, featured above in the first small box. The mean
packet size is 465 bytes. It seems likely that the circled 5 Summary
large packets of nearly 3000 bytes are round initializa- The network traffic generated by Starcraft and
tion or round termination packets, as they strongly Counter-strike look significantly different. There is a
correlate with the end of the decline in packet sizes. significant degree of variation in the packet sizes and
The ovals correlate with large firefights within the bandwidth used in Counter-strike games, in contrast
game. The break in the graph around 1400 seconds, to Starcraft in which the differences between sizes of
indicated by the second large box, is where the map packets sent are barely distinguishable. For larger
was changed to d e a z t e c . games, Starcraft does have more packets transmit-
Figure 6 depicts depicts a CDF of packet interar- ted when the bandwidth requirements increase. The
rivals for Starcraft and Counter-strike. The Starcraft Counter-strike client also follows this model, although
packets arrive fairly uniformly over a range of 10 to the packet sizes are more variable than those in Star-
267cs-assault to de-aztec server bandwidth
0
1 101 201 301 401 501 601 701 801 901 IC01 1101 1M1 1301 1401 1501 1601
Time (seconds)
Figure 3: Counter-strike Bandwidth Sent by Server
.............
0.1 - F
1
Starcraft -
Counter-strike client -------
, ..-..C
Counter-strike server ........
Figure 4: Packet Sizes
268ASIU~ to Aztec Server
3500
.. .A
1.
W
* ~
*: t
c
Figure 5: Sizes of Counter-strike Packets over Time Sent by Server
1 I
1...__._.......~~~-~..
~___,__.._______..__, .' ,____.
,. ......,___.I I........ I
. .
,____--___________1---
...___..
0.9 - .......,
0.8 -
....... 1
0.7 - I
I
,_________-____---.
0.6 -
j
0.5 - : r--'
!
0.4 - I
....... I
________________:
0.3 :
0.2 -
0.1 - Starcraft -
Counter-strike client -------
Counter-strike server ........
o + - ' I I
0 0.05 0.1 0.15 0.2 0.25 3
Figure 6: Packet Interarrival Times
269craft. The Counter-strike server, on the other hand, References
increases the size of the packets when it needs to send [l] Mark Allman. A web server’s view of the transport
more data to the client. These differences are not vis- layer. ACM Computer Communication Review, 30(4),
ible when viewing a bandwidth graph but are impor- October 2000.
tant to note due to their possible effects on network [2] Grenville Armitage. Lag Over 150 Milliseconds is Un-
congestion. acceptable, May 2001.
Starcraft and Counter-strike are also different in [3] Paul Bettner and Mark Terrano. 1500 Archers on a
the bandwidth they consume over time. Starcraft’s 28.8: Network Programming in Age of Empires and
bandwidth consumption varies little over the course Beyond. Gamasutra, March 2001.
of a game regardless of events occurring within the [4]Sally Floyd and Kevin Fall. Promoting the Use
game. Counter-strike, however, has a distinct cyclic of End-to-End Congestion Control in the Internet.
pattern in bandwidth, which vary over time as play- IEEE/ACM Transactions on Networking, February
ers get killed, and have a marked correlation t o game 1999.
events. Overall, the amount of bandwidth consumed [5] Balachander Krishnamurthy and Craig E. Wills. Ana-
by a Starcraft player is comparable to bandwidth con- lyzing Factors that Influence End-to-End Web Perfor-
sumed by a Counter-strike client. A 6-player game of mance. W W W 9 / Computer Networks, 33(1-6):17-32,
Starcraft has each client send between 3000 and 3500 2000.
bytes/ second, and a Counter-strike client connected to [6] Peter Lincroft. The Internet Sucks: Or, What I
a mostly-full server typically sends a little over 3200 Learned Coding X-Wing vs. Tie Fighter. Gamasu-
bytes f second. tra, September 1999.
Counter-strike traffic exhibits a bursty pattern, [7] Dmitri Loguinov and Hayder Radha. Measurement
with 30% of the packets arriving in single bursts. Star- Study of Low-bitrate Internet Video Streaming. In
craft traffic, on the other hand, is much more uniform I n Proceedings of the ACM SIGCOMM Internet Mea-
in its arrival. surement Workshop, November 2001.
Overall, the traffic patterns for Counter-strike and [8] Bruce A. Mah. An Empirical Model of HTTP Net-
Starcraft are optimized for clients with the limiting work Traffic. In Proceedings of IEEE INFOCOM,
bandwidth of a modem connection. Typical Web pages 592-600, April 1997.
browsing and streaming audio can also be accom- [9] S. McCreary and k. clafy. Trends in Wide Area IP
modated by modem connections. However, both Traffic Patterns: A View from Ames Internet Ex-
Counter-strike and Starcraft exhibit different packet change. In Proceedings of I T C Specialist Seminar on
sizes than typical audio or Web packet sizes, with au- Measument and Modeling of I P Trafic, pages 1- 11,
dio traffic having a median packet size of about 300 September 2000.
bytes [lo] and Web traffic having a median packet [lo] Art Mena and John Heidemann. An Empirical Study
size of around 1500 bytes [l]. Moreover, the inter- of Real Audio Traffic. In In Proceedings of the IEEE
arrival times for Counter-strike and Starcraft are sig- Infocom, pages 101 - 110, March 2000.
nificantly different than streaming audio, with audio [ll] Yu-Shen Ng. Designing Fast-Action Games for the
being bursty with a median departure time of about 1 Internet. Gamasutra, September 1997.
second [lo]. These differences present the opportunity
[12] Universiy of California Berkeley. The Network Simu-
to design networks and simulations of networks that
lator - ns-2. Interent site
more effectively support network game traffic above http://uvv.isi.edu/nsnam/ns/.
and beyond the current support for streaming media
and Web traffic. [13] Vern Paxson. End-to-End Internet Packet Dynamics.
IEEE/ACM lPransactions on Networking, Fall 1999.
Notes [14]K. Thompson, G. Miller, and R. Wilder. Wide-Area
The authors would like to thank the anony- Internet Traffic Patterns and Characteristics. IEEE
mous game players who helped obtain the packet Network, pages 10-23, November 1997.
traces. All traces used in providing the analy- [15] Yubing Wang, Mark Claypool, and Zheng Zuo. An
sis for this paper and a Java-based tool t o help Empirical Study of Realvideo Performance Across
parse Commview output can be downloaded from the Internet. In Proceedings of the Ak7M SIGCOMM
http://perform.wpi.edu/downloads/#net-game. The Internet Measurement Workshop, November 2001.
authors would also like to thank the anonymous re-
viewers for their suggestions on how to improve an
earlier draft of this paper.
270You can also read
