Just-In-Time Scaling of Cloud Based Video Games using Machine Learning

Ubisoft’s cloud-based video game ecosystems experience the workload up to 5+ millions players in a typical week. Workloads on game servers are of different scales, ranging from tens of clients per game server to thousands of clients for traditional workloads. To guarantee game player user experience, a pool of servers is launched to react to demands but servers are idles in most of the time. Scaling down servers is even more complex because of the persistent connections to maintain the states and records of players and games. Current scaling tends to guarantee service requirements but incurs significant hosting costs of idle servers. This project aims to investigate a Just-In-Time scaling of game servers. We develop machine learning models to produce more accurate workload in terms of active servers needed. Our approach applies decision fusion to achieve better workload forecasting. Machine learning models are integrated to game platforms of Ubisoft. The success of this research forms a base for a larger scope of collaboration that incorporate more games and online game servers towards a Just-In-Time cost saving scaling system on cloud resources.

Esma Mouine;Jincheng Sun
Faculty Supervisor: 
Yan Liu
Partner University: