Event-based optimization for admission control in distributed service system

In this paper, we consider the problem of admission control in a distributed service system. For the distributed structure of the system, service requests are assigned for service according to the node selection strategy. However, the admitted requests may be dropped since the service capacity of each service node in the system is limited. The desired admission control (AC) policy is to maximize the system performance while keep the dropping probability of the admitted requests below a certain threshold. The event-based optimization framework is applied to formulate the AC process. With the event-based policy, controllers at each service node make decisions only at the epoch when an event of request arrival occurs. The optimization of AC is thus modeled as a probabilistic constrained problem. We present the concept of risk-event to describe the event which happens may lead to dropping of the admitted request, and define the risk of system to denote the probability of a request been dropped. Using the Lagrangian approach, the original problem is converted to an unconstrained problem. Then, by applying the sensitivity analysis method, we derive the performance gradient formula with respect to the policy parameter based on event. Finally, numerical experiments are designed to verify the efficiency of the proposed method.