Game Traffic Analysis: An MMORPG Perspective

1
Game Traffic AnalysisAn MMORPG Perspective
(MMORPG Massive Multiplayer Online Role Playing
Game)

• NOSSDAV 2005

2
Talk Outline

• Motivation
• Trace collection
• Traffic characterization and physical
  explanations
• Conclusion

3
Motivation

• Network games contribute 3 -- 4 Internet
  traffic at backbones in 2000
• MMORPG is very popular, but its traffic is not
  well-studied.
• MMORPG possess some unique features, e.g.,
• no explicit game rounds
• no restrictions on player number
• The objective of playing is not simply to win
  like that in shooting games and war strategy games

4
Popular Genres in MMORPG
Figure curtsey of http//www.mmogchart.com/
5
ShenZhou Online

• a mid-scale, commercial, fatansic MMORPGin
  Taiwan
• thousands of players online anytime
• Technically a typical Asia MMORPG
• Microsoft Windows platform (Windows 2000, MS SQL
  Server)
• Data communication based on TCP

6
?? Online
Its me
7
Trace Collection
a FreeBSD PC w/ tcpdump
20 hours, 110,000 connections 1,356 million
packets in total
8
Trace Collection - Summary
Trace Sets Date Time Period Drops
N1 3 8/29/04 (Sun.) 1500 8 hr. 0.003
N2 2 8/30/04 (Mon.) 1300 12 hr. ?
Trace TCPConn. Packets (in/out/both) Bytes (in/out/both)
N1 57,945 342M / 353M / 695M 4.7TB / 27.3TB / 32.0TB
N2 54,424 325M / 336M / 661M 4.7TB / 21.7TB / 26.5TB
1,356 million packets in total
9
Traffic Characterization

• Packet size
• Bandwidth usage
• Within-connection traffic patterns
• Aggregate traffic patterns

10
Payload size distribution - CDF
Tiny packets
98 client payload lt 32 bytes
Overall avg. pkt size 84 bytes, contrast to 400 bytes in backbones
Protocol overhead
73 bytes are headers
30 bytes are pure TCP acks
Dominant commands
In client traffic, two payload size occupy 90
11
Packet load distribution
Data packets Data packets
Server lt 5 pkt/s
Client lt 15 pkt/s
Ack packets incl.
Server Client lt 15 pkt/s
12
Bandwidth Usage
Data packets Data packets
Server lt 10 Kbps
Client lt 3 Kbps
Ack packets incl. Ack packets incl.
Server lt 13 Kbps
Client lt 8 Kbps
A Comparison
ltlt 40 Kbps than a first person shooting game, Counter-Strike
Similar to Warcraft III, a realtime strategy game
13
Cannot Overlook MMORPG Traffic

• Ragnarok Online in Taiwan announced a record of
  370,000 concurrent players
• assuming each player uses 10 Kbps ?total 3.7
  Gbps is required for only ONE game

14
CDF of Server Packet Interarrival Times
50 are around 200 ms ? evidence of periodic
message dispatch
15
CDF of Client Packet Interarrival Times
deviated from exponential distribution
16
?? Online
These players are idle and keeping stalls.
17
Diversity in User Behaviors

• active player fits to exponential
• idle player uniform determinist distr.
• regular player in-between the two extremes

regular player
idle player
18
Temporal Locality in Client Traffic
Auto-correlation function
clustering nature of player actions clustering nature of player actions
Active actions walking, fighting
Inactive actions talking, viewing equipments, trading, idle
19
Clustered user actions
Client packet inter-arrival times from a certain
player
combat
combat
walk
20
Temporal Locality in Server Traffic
Auto-correlation function
1. Server pkt primarily convey position updates
2. The rate of position updates depends on the number of neighbors (characters within scope) 3. The spatial locality in of neighbors can be transformed to the temporal locality
21
Spatial Locality in of Neighbors
Pos Neighbor
(3,3) 1
(2,3) 2

(1,1) 4
(1,2) 5
(1,3) 4
22
Spatial Locality can transform to
TemporalLocality
the spatial locality shows up in terms of
temporal locality in the traffic as a characters
move continuously on the map
t Neighbor
1 1
2 2
3 1

10 4
11 5
12 4
23
Temporal locality in server traffic
Server packet rates for a player
In busy place, or w/ rapid actions
mean
Idle,isolated
24
Aggregate Packet Arrivals

• Aggregate packet arrival time series
• count the number of incoming/outgoing packets
  regardless of connections
• take samples every 10 ms

25
ACF of aggregate server pkt arrivals
(taken in every 10 ms)
Auto-correlation function

• periodity 200 ms
• position updates are synchronous for all clients
• incurring periodic packet bursts

26
ACF of aggregate client pkt arrivals
Positive temporal dependence up tothree minutes.
27
The flash crowds effects
Auto-correlation functions

• ACF for successive 3 minutes
• Players are active simultaneously at times
• global eventslead to flash crowds effects

28
Frequency Components in Server Traffic
Frequency ident.
multiples of 5 Hz
frequency seems to be adjusted by some metrics, e.g. of neighbors
Power spectral density function
5 Hz
Multiples of 5 Hz
29
Frequency Components in Client Traffic
Power spectral density function
Frequency ident.
multiples of 6 Hz
by auto-walk and auto-attack timers
freq. is adjusted by level / skill / weapon held
6 Hz
Multiples of 6 Hz
client timers are synchronized ? packet bursts
30
Conclusion

• Traffic characteristics physical explanations
• Tiny packets(TCP maybe an overkill for MMORPG)
• Low bandwidth requirement for each client
• Non-poisson user inputs (diversity of user
  behaviors gt no typical user gt difficult to
  provide a general source traffic model)
• Temporal dependency in within-connection traffic
  (clustering nature in user inputs)
• Temporal dependency in aggregate traffic (flash
  crowds effects)
• Periodicity and synchronization(common
  implementation practice)

31
Questions?
32
Thank You!
33
Backup Slides
34
Thank You!
Kuan-Ta Chen http//kuan.ilife.cx/
35
Network Games in Taiwan

• 120 network games in 1st quarter, 2003
• market value of 6,800 million NT dollars in 2003
  ? 80 share of the whole PC game industry
• Gamania (Lineage) owns gt 4,000 Mbps links
• R. O. has a record of 370,000 players online,
  equivalent to 1.5 of the population

36
The Increasing Popularity of MMOG
Figure curtsey of http//www.mmogchart.com/
37
Acknowledgement

• Tsing-San Cheng, Larwence Ho, Yen-Shuo Su,
  Chen-Hsi Li

UserJoy Technology, Co., Ltd.
E2 Technology, Co., Ltd.