Study of oscillations in admission control algorithm for web servers

In recent years, the advance of the Internet technology and e-commence applications becomes the motivation for the development of scalable server brokerage architectures for the purpose of load sharing. There are many ongoing researches and proposed solutions in solving this problem, but our project focus on providing the satisfactory QoS with the brokerage architecture. Besides the load sharing among all the servers in the server pool the brokerage architecture has to guarantee the response time provided by the web server since it is an important factor for user satisfaction. But in the earlier server selection algorithm of the brokerage architecture, a threshold is used to decide whether to accept/reject users depending on the current response time. A problem with this approach is that oscillations occur. Due to the abrupt manner of accepting/rejecting user, the system experiences unavoidable oscillations in terms of the response time, the number of users in the system and even the utilization of the servers. The work of this thesis is oriented towards solving the problem of system performance oscillations. We first establish the relationship between the number of users in the system and the average response time. Then we study a theoretical model of the oscillation of the number of users in the system. Then we propose a probabilistic approach to admission control where the probability of rejecting a new user increases as the load increases. Using the theoretical model, we prove that with a probabilistic approach, the oscillations will normally be suppressed, and the number of users in the system reaches a stable point. We 3 also test the effect of different probability functions and the impact of different inter-observation time intervals on the oscillation by careful simulation experiment. Finally, the probabilistic approach is used to provide differentiated classes of service to different user groups. We show that the probabilistic approach provides better performance than the on-off decision approach.