Network Game Design: Hints and Implications of Player Interaction

After the relationship between each pair of players has been determined, we divide players into clusters: players who are neighbors (must be mutually) form a horde, while players who are partners (must be mutually) form a group. Note that a player may simultaneously belong to a number of groups, because s/he can form a group with some players and form another group with some other players in different periods. Finally, hordes containing the same players are merged into a place. This operation forms disjoint places, where each player uniquely belongs to a place. To demonstrate the principle behind the identification algorithm, in Fig. 4, we depict the packet arrival processes of a four-player group. We can see that the fluctuations of server packet arrivals follow very similar patterns during the co-presence time, which is longer than 40 minutes. Meanwhile, the client packet arrivals exhibit similar patterns in their activity levels. The figure also shows that the duration of a group may be longer than the session time of any member. In other words, even the members in a group come and go, a group remains if at least two members are still in the game.

5  Analysis of Player Interaction and Its Implications

In this section, we analyze players' interaction patterns and their implications for network game design. We first present a summary of the group structure, which is inferred using the Algorithm 1 in Section 4. We then analyze player interaction patterns and discuss their implications in a number of aspects, including player density, network proximity, session duration, and interaction levels.

5.1  Basic Statistics

Of the 16,528 players who had potential to interact with others, i.e., players with session time longer than ten minutes, 67% were social players, and 17% were group players. On average, a social player interacted for 70% of his online time, while a group player teamed up with others for 50% of his/her online time. Meanwhile, a social player had 3.5 neighbors and a group player had 1.4 partners on average, as shown in Table 2. Fig. 5 shows how players were scattered in different places and groups. We can see that 40% had at least four neighbors during the game, while 20% stayed in places where more than one hundred players had gathered. The most popular place, where 1,816 players gathered, shows that more than 10% of players spent the whole session in the same location. On the other hand, groups are frequent but mostly small-sized. Among 1,389 groups, 85% were "duos," and only 2% had more than four members. We believe the prevalence of "duos" is not unique to this game, as it reflects the fact that some players are not really social and usually interact primarily with one other person in real life. This echoes the following comments by gamers reported in [15]:

"Distinct from `grouping' in that [sic] (as I see it), the `duo-group' is most often the same two people (perhaps with a relationship offline), who don't typically group with people other than each other."

"My answers to questions in this vein would show I often group, but it [sic] would likely be grouped with others who are honestly much more social than I, since I predominantly duo with my spouse."

5.2  The Skewness of Place Populations

As the number of concurrent players in a MMOG are usually exceeding tens of thousands, even a powerful server could not handle the vast computational requirement of an entire game. A straightforward solution for solving the scalability problem is to distribute the workload among multiple servers. There are two common approaches: the server-cluster approach [1,20], and the peer-to-peer approach [10], both of which normally divide the game world into a number of regions with each of which may be managed by different servers. With this approach, the key to an efficient load balancing depends on whether it can correctly delegate a region (of the game world) to an appropriate server, while minimizing the overall transmission latency and bandwidth utilization and subject to the constraint that no server is overloaded. Furthermore, one of the deciding factors that affects the outcome of region delegation is how players are distributed across the virtual world. If players tend to be uniformly distributed, then a straightforward fixed-size partitioning of the game world is sufficient, given the players in each region is not too many to overload any single server. On the contrary, if players tend to aggregate in certain hotspots that change from time to time, then a dynamic and adaptive partitioning algorithm must be used. From the distribution of place popularity (Fig. 5), there were a number of crowded places, e.g., places where more than fifty players gathered, which indicate the existence of hotspots. Further analysis shows that the distribution of populations in different places can be well modeled as the concatenation of two Zipf-like distributions [19]. Suppose a place with rank i and population p_i follows a Zipf-like distribution; then p_i=C/i^α, where C is a normalizing constant and α > 0 is the power-law exponent, i.e., the population of a place is inversely proportional to its rank with a log-log transform. The smaller the α, the heavier tail the distribution will have. Fig. 6 depicts the populations of places versus their ranks on a log-log plot, where the points are scattered along two line segments. This evidences that the populations of the most popular twenty places follows a Zipf-like distribution with α ≈ 1.4, while the populations of the less popular places follows the same distribution with α ≈ 0.5. Even the populations of popular places have a lighter tail than a standard Zipf distribution (α = 1), the most crowded hotspot has about 10% of the players aggregated, and the most popular 1% of places (out of 7,712) contain about 30% of the players. This manifests that the dispersion of players is far from uniform and is closer to a power-law distribution. The high variability in player density across a virtual world implies that static world partitioning strategies [10,18] are inadequate for MMOGs. On the other hand, we remark that dynamic partitioning algorithms, which adaptively adjust the number, size, and ownership of regions based on the constantly changing player density, should be employed for online games with either server-cluster [14,2] or peer-to-peer architecture [7,11].

5.3  Network Proximity of Interacting Players

A number of peer-to-peer infrastructures for online games and network virtual environments (NVEs) have been proposed in recent years. However, a fundamental design question remains unsolved: How should the participating hosts interconnect for better networking performance? Currently, there are three common approaches for constructing overlay networks: 1) numeric-ID based [7]; 2) network topology-based [6], e.g., optimizing network latency; and 3) in-game location-based [18,8]. Although the first two approaches are adequate for general purpose applications, such as file sharing, they may be less appropriate for NVEs. For example, with a latency-optimized overlay, a message sent from a host can be delivered to any other host in the network with generally low latencies. However, if game players only communicate with their neighbors in the virtual world, then a naive approach, where the hosts of each pair neighbor are interconnected, would achieve better performance in terms of message transmission latency. A complete evaluation of the design alternatives (i.e., different strategies to connect overlay nodes) requires a detailed trace involving inter-host measurements and is beyond the scope of this study. Instead, using a comparative approach, we investigate the degree of correspondence between the locality in the network and the locality in the virtual world, which implies the similarity of the overlay networks built with the second approach and the third approach.

5.3.1  Similarity in Network Addresses

Fig. 7 shows the proportion that a pair of players with a certain kind of relationship share the same IP address prefix. We found that about 7% of partner-pairs co-located in the same /16 network, whereas only 1.7% of unrelated-pairs were in the same network. At the same time, approximately 5% of partner-pairs used the same IP address, while very few unrelated-pairs did so. This indicates that a proportion of group members played at the same physical location (in an apartment or Internet café) through a NAT device. Comparing the three levels of relationship, we can see that players with a higher level of interaction tend to be closer in network topology. These findings suggest that overlays built on in-game localities correlate with topology-aware overlays to some extent, i.e., whether the overlay is based on in-game closeness or network topology does not make so much difference.

5.3.2  Similarity in Network Latencies

Fig. 8 provides another view to examine the network proximity of players. We use the difference in minimum RTTs (round-trip times) and the difference in average RTTs between the players' hosts to gauge their network distance from the perspective of game servers. The figure shows that interacting players are much closer to each other compared with unrelated players. This echoes the agreement between in-game locality and network locality based on IP addresses.

5.3.3  Advantages of the Property

Client-server based games would benefit from the network proximity property of interacting players in terms of proxy placement and fairness of game playing. A client-server architecture with proxy support is a design that alleviates the single bottleneck problem of, and retains the manageability of, the centralized server architecture [12]. In such architecture, players connect to their nearest proxy server, which processes and responds to user commands in place of game servers. Proxy servers are usually close to users; thus, the transmission latency and the user-perceived response time could be significantly reduced. However, additional proxy-to-proxy communication is required when players who connect to different proxy servers interact. In this case, the overall message delivery latency could be longer due to detours between proxys. To provide high responsiveness and interactivity, two seemingly contradictory conditions need to be met: 1) players should connect to their nearest proxy server, and 2) players who are neighbors in the game should connect to the same proxy server. Fortunately, the network proximity property of interacting players resolves the contradiction between the above conditions. The reason is that players are likely to connect to the same proxy as their in-game neighbors (with a non-significant probability), even if proxy selection is only based on the network distance. Fairness is also an important issue in network game design. Normally there is a bias against players with longer latencies (between clients and servers), since their screens tend to update later than others if the server releases an update message for all participants at the same time. Furthermore, the actions of players who are far from servers are likely be processed (by servers) later than others, even if all commands are issued simultaneously (on their respective computers). In the case of client-server based games, a set of players can have fairer play if their network latencies to game servers are similar. It is because that servers will process their actions approximately at the same time if they trigger commands simultaneously. Thus, the network proximity for interacting players makes their interaction, such as fighting each other, fairer, as their network latencies to game servers are similar. For the above reasons, we consider that the network proximity of interacting players is a "benign" property, as it is favorable to the design of network games with both peer-to-peer and client-server architectures. We remark that as player interactions are generally spatially localized, the mapping of spatial locality between the game world and the network topology would be rather important and could be a key to the optimization of network latency in online games.

5.4  Game Playing Time and Social Interaction

As social interaction is one of main attractions of MMORPGs [17], a factor that encourages time investment and personal attachment, one may ask: Does social interaction "tie" players to a game and encourage them to play longer than independent players who do not have any interaction with others? To answer this question, we plot the survival curves [9] for sessions grouped by interaction levels, and check the significance of the differences between the groups. In Fig. 9, the survival curves for three groups of sessions are plotted. Visually, the curves diverge significantly from each other; and the log-rank test reports p=1−Pr_χ²,2(1532) ≈ 0, which indicates the sessions in different groups are far from equivalent. The median session time for independent players, social (but solo) players, and group players is 40, 112, and 190 minutes, respectively, which gives a high ratio of nearly 5:1 for group players and independent players. This figure implies that there is a strong relationship between social interaction and time investment in a game. Game designers, therefore, may provide more facilities to encourage teamwork, or even force players, to form groups for quests, which would increase the "stickiness" of a game and eventually generate more revenue through subscription fees.

5.5  Degree of Group Interaction

Now that the existence of network proximity and the tendency of group players to be "tied" in a game have been established, in this subsection, we focus on group players and investigate the relationship between the level of group interaction and some potentially relevant factors. We will show that the former is related to the degree of network proximity and group size. To quantify the level of group interaction, we define the "group time" of a group as the average pairwise co-presence time among its members. We plot the relationship between group time and three potentially relevant factors in Fig. 10. As can be seen in Fig. 10(a), the closer group members were in network topology, the longer they played together. The property of network proximity can also be examined in terms of the network latency between clients and servers. The smaller the difference in minimum RTT experienced by group members, the longer the group playing time they shared, as shown in Fig. 10(b). Assuming that network proximity reflects geographical locality, the above relation implies that 1) real-life relationships carry over into the virtual world, and/or 2) real-life interaction plays a key role in game playing. The latter explanation suggests that enriching in-game communication supports, e.g., audio and video chat facilities, would encourage players to be more involved in team cooperation. Fig. 10(c) shows that players in larger groups tend to stay longer. This phenomenon can be explained by the enjoyment derived from interaction and social bonds. For example, the latter may prevent a player leaving a game prematurely, as it would affect the group's adventure. This suggests that encouraging players to form larger groups would also increase the "stickiness" of the game. However, the effect diminishes for groups with more than five members, possibly because the role of each player in a large group is less important. In other words, the social bond is weak when a player leaves a large group than leaving a smaller group.

6  Conclusion

In this paper, we have analyzed player interaction for ShenZhou Online, a mid-sized commercial MMORPG, and drawn out a number of hints and implications for network game design. The analysis reveals that the dispersion of players in a virtual world is heavy-tailed, which implies that static and fixed-size partitioning of game worlds is inadequate. We have shown that neighbors and teammates tend to be closer to each other in terms of network topology, a property that is beneficial to both client-server and peer-to-peer based games, because message delivery between the hosts of interacting players is faster. In addition, this property makes games fairer, as interacting players tend to have similar latencies to their servers. We have also found that participants who have a higher degree of social interaction tend to play much longer than independent or solo players. This suggests that game companies could increase the "stickiness" of games by encouraging, or even forcing, team playing. Furthermore, the duration of group play correlates with the size of the group and the network distance between the players. Specifically, players who are closer in network topology tend to team up for longer periods. Larger groups generally lead to longer collaboration due to the enjoyment derived from interaction and social bonds, which suggests that games could also be made stickier by encouraging the formation of such groups. We believe that only through observation can we fully understand what users desire and therefore be able to design appropriate network systems for them. For instance, via an inspection of how players react to network conditions, we can understand how players perceive unfavorable network QoS (e.g. packet loss and delay jitters are more intolerable than network latency). Such understanding could be very helpful in the design of network game systems, especially in the message delivery subsystem, to diminish the impact of unavoidable network impairment on players as possible [4]. Designing more efficient network gaming systems which take account of user behavior and perception and be optimized for user satisfaction will be one main theme of our future work.

References

---

Sheng-Wei Chen (also known as Kuan-Ta Chen)
http://www.iis.sinica.edu.tw/~swc 
Last Update September 28, 2019