Session-Level Load Balancing for High-Dimensional Systems

Load balancing is critical for the performance of big server clusters. Although many load balancers are available for improving performance in parallel applications, the load-balancing problem is not fully solved yet. Recent advances in security and architecture design advocate load balancing on a session level. However, due to the high dimensionality of session-level load balancing, little attention has been paid to this new problem. In this paper, we formulate the session-level load-balancing problem as a Markov decision problem. Then, we use approximate dynamic programming to obtain approximate load-balancing policies that are scalable with the problem instance. Extensive numerical experiments show that the policies have nearly optimal performance.