A success-history based learning procedure to optimize server throughput in large distributed control systems

Abstract

Large distributed control systems typically can be modeled by a hierarchical structure with two physical layers: Console Level Computers (CLCs) and Front End Computers (FECs). The control system of the Relativistic Heavy Ion Collider (RHIC) at Brookhaven consists of more than 500 FECs, each acting as a server providing services to a potentially unlimited number of clients. This can lead to a bottleneck in the system, as heavy traffic can slow down or even crash a system, making it momentarily unresponsive. In this paper, we consider this problem from a game theory perspective. Specifically, we consider the case where the server has a varying capacity. First, we model this problem as an integer programming problem. Second, we adopt a regret-based procedure as a basic solution and then propose a success-history based scheme to better accommodate the dynamic server capacity. Finally, simulation results show that both algorithms perform well and lead to a significant improvement of system performance. Moreover, compared with the regret-based procedure, the proposed success-history based scheme results in a higher server throughput and lower crash probability under the dynamic environment.