Multimedia & Communications Lab
Seoul National University
Multimedia & Communications Lab
Seoul National University
Multimedia & Communications Lab
This paper describes the measurement and analysis of a Massively Multiplayer On-line Role Playing Game(MMORPG). About 3 billion packets were traced for around 8 days. The results showed that MMORPG traffic has very small packet size and bursty inter-arrival time. By the analysis of per-flow inter-arrival time and average connection time per flow, we showed the gamers’ behavior pattern. Tendency of bandwidth consumption during the measurement period suggested a hint for the game server management. Lastly, the correlation between number of users and bandwidth showed linearity and this fact can be an important reference information for anticipating future demands of game servers and networks. Through the analysis of traced packets we could build traffic models of MMORPG, especially for packet size and inter-arrival time. The model can be used for game traffic generator and for network simulations
MMORPG(Massively Multiplayer Online Roleplaying Game), traffic, inter-arrival time, traffic model.
With the advances of Internet infra structure, ratios of traffics are also changing dynamically. Recently, Online Game Traffics are growing rapidly according to their popularity and provisioning. There are many kinds of Online Games. But in the aspect of traffic measurement, the games of large bandwidth consumption and large participants arouse our interests. MMORPG(Massively Multiplayer Online Role Playing Game)s are those.
Measurement of the Internet traffics has been done on many sorts of applications and some kinds of online games has been measured. But intensive measurement and analysis of MMORPG traffic has not been conducted yet. Common characteristics of online game traffics are small and highly periodic UDP packets. But most MMORPGs uses TCP packets because of client server structure and connection management convenience.
World 1st MMORPG and world largest MMORPG are serviced in Korea. ‘Lineage’ is the latter one. Simultaneous participants in this game has exceeded 300 thousands. It has over 2 million registered users around the world. We have captured ‘Lineage’ traffic for 8 days and could store 281 Gbytes of raw data.
In section 2, the details of measured results and their analysis are presented. In section 3, the traffic model for MMORPG is also suggested.
2.CHARACTERIZATION OF MMORPG TRAFFIC
To measure the MMORPG traffic, we used ‘tcpdump’. This tool shows the timestamp, IP address, port number, data size, TCP flag and so on. Our measuring machine was operated on LINUX. Measuring machine was connected to the optical gigabit-switch where the game server was connected and we used port mirroring function to minimize additional traffic load on the network.
Table 1 shows the statistics about overall packet number and packet size. Server packets are larger than the client packets because they contain data of multiple clients. The smallest packets have no data and have just 40 bytes of header. They are pure control packets such as SYN, ACK and FIN.
Figure 1 Distribution of Server Packet Size
Figure 1 shows the distribution of server packet size and Figure 2 shows cumulative distribution of server packet size. Hereafter, packet size represents the pure data bytes excluding the header bytes. Both figures are about server packets. As shown in the figures, packet sizes are very small and are narrowly distributed. This characteristic is similar with the results of counter-striker game. Even though the average packet sizes of MMORPG and non-MMORPG are different, common point is their packet sizes are very small. Distribution of client packet size also shows the similar tendency in Figure 3 and Figure 4.
Figure 2 Cumulative Distribution of Server Packet Size
Figure 3 Distribution of Client Packet Size
Figure 4 Cumulative Distribution of Client Packet Size
2.3Inter-arrival & Inter-departure time
We analyzed the inter-arrival time of packets to server first. Maximum inter-arrival time was 20 seconds and the minimum was 0 second. Average value was 386 usec. In figure 5, we can see that 90% of packets inter-arrive within 0.8 msec. and 99% within 2 msec.
Inter-departure time from server has also similar values with inter-arrival time. The average value is 438 usec. This is less frequent than inter-arrival time. That means clients send packets more frequently than server does. Figure 6 shows that 88% of packets departed within 1msecs. and 99% within 4 msecs.
Figure 6 Cumulative Inter-departure Time
2.4Inter-arrival time per flow
In the previous section, we showed the inter-arrival time of overall traffics. Here, we present one more inter-arrival time. That is per-flow inter-arrival time. Flow means packet streams that come and go from one specific client to the server. Per-flow inter-arrival time designates a gamer’s behavior, especially his or her action interval during the game. Figure 7 and 8 shows distribution of per-flow inter-arrival time and cumulative distribution of it, respectively.
Figure 7 Distribution of per-flow inter-arrival time
Figure 8 Cumulative Distribution of per-flow inter-arrival time
Peak point of per-flow inter-arrival time stood at 200ms. 98% of per-flow inter-arrival time was shorter than 800ms. It implies that a gamer sends at least a packet in a second. Average per-flow inter-arrival time was 263.58 msec and it means that a gamer sends approximately 4 packets per second to the server.
2.5Average Connection Time Per Flow
Among 215,254 total flows, 161,648 flows lasted more than or equal to 1 second. We counted these ones for the calculation of average connection time because flows with shorter connection time were not really connected to the server. Average connection time of these really connected flows was 2980.846 seconds. This shows that average playing time of a gamer is about 49.68 minutes.
We analyzed the tendency of bandwidth consumption during the measurement period. Figure 9 shows the periodicity of bandwidth consumption. Between 11 a.m. and 10 p.m. every day, bandwidth was consumed more than in another durations. Around 6:30 a.m., traffic was least. Peak points were found on holidays. 11th Aughust was Sunday and 15th August was Korea Independent Day. In the afternoon(2 ~ 4 p.m.) on those days, bandwidth consumption curve was on its summit.
Bandwidth drop points where the bandwidth falls near the bottom are made by periodic system check and system anomalies.
Figure 9 Bandwidth Consumption During a Week
2.7Correlation between number of users and bandwidth
Different on-line games have different correlation between number of users and bandwidth according to their game structures. In MMORPG like ‘Lineage’, what does the number of users have to do with bandwidth? Just linear relation? Or exponential relation? To clarify the correlation between them, we separated all the data into segments of 1 minute duration. The number of users and their traffic volume was calculated for each 1 minute segment. Thus, we could get 11983 segments.
Figure 10 shows the result. Here, we could find out that the correlation between number of users and bandwidth is linear. Some can argue about the result because when the number of users increases, the interaction between game server and clients can increase exponentially. For example, if there are 5 clients connected to a server and a client is added, then a message from a client should be broadcast to 5 other clients. In this way, if each of 6 clients sends a message to the server at the same time, then 30(6*5) messages should be broadcast to every other clients. But in case of ‘Lineage’, not every client’s message to the server is broadcast. A client’s message is forwarded just to a small group of other clients who are in the same area. So, the effect of user increase is restricted to a small amount. This result can be a reference information when the network and server of service company is redesigned.
Figure 10 Correlation between Number of Users and Bandwidth
When we simulate network traffic, the first step is to make traffic model. The traffic model needs two submodels. Packet size and Inter-arrival time are them. Each of them has different characteristics so the models of them are inevitably different. There are several mathematic models that can be adopted as the traffic model of MMORPG. Among them two models matched best the real traffic.
The distribution of packet size best matches ‘Power Lognormal Distribution’. Following equation is the distribution function of ‘Power Lognormal Distribution’.
Here, is probability density function of standard normal distribution and is the cumulative distribution function of standard normal distribution.
Cumulative distribution function of ‘Power Lognormal Distribution’ is as follows.
From two equations above, by adjusting and values, we could get the proper distribution curve similar to the one we measured.
Inter-arrival time doesn’t match any of existing mathematic models. But the most appropriate model was ‘Extreme Value Distribution’.
Here, is location parameter and is scale parameter. Cumulative distribution function of ‘Extreme Value Distribution’ is as follows.
On-line games in the Internet are getting more and more popularity. Due to this tendency, bigger traffic loads are imposed on the Internet. Among those on-line games, MMORPG has special characteristics in the aspect of the size of simultaneous participants and its traffic burstiness. So, we measured an MMORPG and analyzed its characteristics and made an traffic model for MMORPG.
First we analyzed the distribution of packet size and the overall inter-arrival time and inter-arrival time per flow as well. Average connection time per flow was analyzed to show the trend of gamers’ behavior. For the convenience of service company’s management affairs, tendency of bandwidth consumption during a week was analyzed. Correlation between number of users and bandwidth can be the reference information for future restructuring of server and network facilities.
Based on the above analyses, we could propose the traffic model for MMORPG. Packet size distribution could be modeled by ‘Power Lognormal Distribution’ and inter-arrival time by ‘Extreme Value Distribution’
This research result is restricted in its generality because just one game was measured. For the improvements of its generality, some more measurements of other on-line games are needed.
Yu-Shen, Designing Fast-Action Games For The Internet, Gamasutra, Sep. 5, 1997