Massively multiplayer role-playing games (MMORPGs) are among the most popular types of online game. A successful title may have tens of thousands or even millions of subscribers and, at any given time, may have thousands of players online. This paper presents a method for modeling MMORPG system performance and applies it in an analysis of two real MMORPGs. The results show that a strong linear relationship exists between performance metrics at the server side and the number of concurrent players online. As a result, utilization of IT resources, including network traffic and server load, can be predicted, given the number of concurrent players. The performance model presented here can be used for automated IT resource allocation at runtime and is thus useful in the context of utility computing and on demand systems.

Online games are the trend of the day. Online game services such as PlayStation ** online, Xbox ** Live, GameSpy ** Arcade, and other independent PC-based services are becoming increasingly popular. In their 2004 prediction, International Data Corporation (IDC) estimated industry revenue to reach $656.3 million in 2004 and grow to over $2 billion by 2008. (1) Online games enable multiple players to simultaneously interact in a "game world" to which they connect over a network. Most online games today follow a client/server model; research is ongoing concerning the feasibility of other types of architecture, such as peer-to-peer and grid architectures. (2,3)

PC-based online games can be classified into two subcategories--multiplayer online games (MOGs) and massively multiplayer online games (MMOGs)--based on the maximum number of simultaneous players in a single game world. Popular MMOGs might have thousands of players online at any given time, usually exclusively on a company-owned server. On the other hand, MOGs usually have less than 50 players online and are usually played on private servers. MOGs are frequently adopted by fast-paced game genres like first-person-shooter (FPS) games in which response latency is the most critical factor, aside from game content, in the game experience. Some examples of FPS games are Doom, ** Quake, ** Half-Life, ** and Counter-Strike, (4,5) which are session-based games in which the goal is for a player's alter ego to accumulate successful "kills" against other players. To support more players, MOGs can scale up by horizontally replicating the game world without coordination or synchronization between these worlds.

The first and most popular type of MMOG is the massively multiplayer online role-playing game (MMORPG) genre, which can be traced to the nongraphical online multiuser dungeon (MUD) games of the 1970s and became popular in the late 1990s. Reference 6 estimates that MMORPGs hold a 95.5 percent share of the MMOG market. Some examples of MMORPGs are EverQuest **, Lineage **, and World of Warcraft **. (7-9) MMORPGs are also called persistent-state-world (PSW) or persistent-world (PW) games because the game world is normally hosted by a company and is always available, and world events happen continually, even while some of the players are not playing their character. Players may retain the same title for several years. Popular MMORPG game titles have large numbers of subscribers. Figure 1, based on Reference 6, shows the statistics up to May 2005.

From the figure, we can see that subscriptions to the same title vary over time. Such dynamics of subscription bring challenges to gaming service providers who traditionally install a dedicated infrastructure for each title, due to the high risk of over- and under-allocation of resources and potentially poor resource utilization. This situation may become worse, as the number of commercially operated game titles has increased dramatically in recent years.

A successful gaming service provider must be able not only to satisfy its subscribers' demand for high quality and attractive game content but also to reduce the risk of high investment in game hosting infrastructure associated with the difficulty of predicting the success of a new game title. The utility computing model, (also called "on demand," (10) "utility data center," (11) or "just in time computing" (12)) is believed to be the solution to this problem from an infrastructure perspective. For example, Shaikh et al. propose an on demand service platform for hosting large-scale multiplayer games. (13) The key idea proposed is sharing IT resources across multiple game titles or customers by dynamically provisioning and deprovisioning resources for a title or customer from a shared resource pool.

A critical component for these solutions is the provisioning manager, which is responsible for resource provisioning and deprovisioning, such as the IBM Tivoli Intelligent Orchestrator (TIO). The TIO is an off-the-shelf product that automatically deploys and configures servers, software, and network devices in a data center environment. (14) The primary function of a provisioning manager is to collect performance and availability metrics from game servers, predict their trends, and decide how to adjust resource allocation accordingly