A Consistency Optimization Framework for Multi-server Distributed Interactive Applications

Distributed interactive applications (DIAs) refer to Internet-based applications that emphasize on real-time interactions. They are becoming very popular in recent years due to some successful applications, such as online games, distributed virtual environments and collaborative design systems. As these applications are accessible to any users connected to the Internet, they are faced with a critical problem — how do they maintain their service level as the number of users increases? Unlike the scalability of non-real-time Internet applications, DIAs have a much tighter time constraint to meet as users expect real-time interactions and feedbacks from the applications.One of our research focuses in the past decade has been trying to address this scalability problem of DIAs based on multi-server load balancing techniques. In our earlier works, we have developed efficient load balancing methods aiming at quickly resolving the server overloading problem. Although we were able to propose very efficient methods to address the problem, we notice that even though we may resolve the server overloading problem efficiently, the load balancing process itself of all existing load balancing methods introduces additional delay, including processing and extra server-server delays. These delays lead to two problems, the consistency problem due to user interacting with each other and the accuracy problem of the load balancing solutions due to the drift in server loading.In this project, we propose to investigate these two problems. To address the consistency problem, we propose to develop an optimization to study the trade-offs between the effective resolution of inconsistencies and the load balancing of users among multiple servers. With a goal of improving the quality of service and users’ experience of DIAs, we plan to find an optimal way to partition workload so as to minimize inconsistency due to network latency, while ensuring that the loads of all servers are below a given threshold. We achieve this objective through classifying user movement patterns, quantifying the amount of inconsistencies of each pattern, and formulating an optimization to partition the workload in such a way that interacting users would be served by the same server as much as possible to minimize inconsistency. To address the accuracy problem of load balancing solutions, we propose to develop a delay adjustment method to compensate for the drift of server loading due to processing and network delays. Our approach to these two problems is unique and has not been addressed in the context of load balancing in DIAs. This project will allow us to look at the load balancing problem of DIAs from a different angle and is expected to bring significant impact to the design of multi-server DIAs in the long-run.