Massively Multiplayer Online Games (MMOGs) are a very popular class of distributed systems with more than 20 millions of active users worldwide. MMOG have strong applicative requirements in terms of data consistency, persistence, responsiveness and scalability. Shaped by the behavior of the players in-game, MMOG workloads are data-intensive and hardly predictable. Despite extensive research in the area, none of the currently existing architectures is able to fully satisfy all the requirements in presence of such complex workloads. This thesis investigates the ability of MMOG architectures to better accomodate the workload by monitoring player activity at runtime. By doing that, the system is able to detect evolutions that are hard to foresee at startup, and dynamically allocate ressources to handle the load.We describe different techniques of runtime player monitoring and propose mechanisms to incorporate user behavior in the architectural design of MMOGs. Our experimentations are based on realistic workloads and show that our mechanisms have negligible overhead and improve global performances of MMOG distributed architectures.
