Performance evaluation of a measurement-based algorithm for distributed admission control in a DiffServ framework

Distributed admission control solutions share the idea that no coordination among network routers (ie, explicit signaling) is necessary, when the decision whether to admit or reject a new offered flow is pushed to the edge of the IP network. Proposed solutions differ in the degree of complexity required in internal network routers, and result in a different robustness and effectiveness in controlling the accepted traffic. This paper builds on a recently proposed distributed admission control solution, called GRIP (Gauge and Gate Reservation with Independent Probing), designed to integrate the flexibility and scalability advantages of a fully distributed operation with the performance effectiveness of admission control mechanisms based on traffic measurements. We show that, in the assumption that traffic sources are dual-leaky-bucket shaped, GRIP allows providing deterministic performance guarantees. Tight QoS performances are made possible even in impulsive load conditions (i.e., sudden activation of several flows), thanks to the introduction of a "stack" mechanism in each network node. A thorough performance evaluation of the conservative effects of the stack shows that the throughput reduction brought about by this mechanism is tolerable, and limited to about 15%