Disturbance Rejection and Control in Web Servers

Abstract: An important factor for a user of web sites on the Internet is the duration of time between the request of a web page until an answer has been returned. If this response time is too long, the user is likely to abandon the web site and search for other providers of the service. To avoid this loss of users, it is important for the web site operator to assure that users are treated sufficiently fast. On the other hand, it is also important to minimize the effort to optimize profit. As these objectives often are contradictory, an acceptable target response-time that can be formulated. The resources are allocated in a manner that ensures that long response times do not occur, while, at the same time, using as little resources as possible to not overprovision. The work presented in this doctoral thesis takes a control-theoretic perspective to solve this problem. The resources are considered as the control input, and the response time as the main output. Several disturbances affect the system, such as the arrival rate of requests to the web site. A testbed was designed to allow repeatable experiments with different controller implementations. A server was instrumented with sensors and actuators to handle requests from 12 client computers with capability for changing work loads. On the theoretical side, a model of a web server is presented in this thesis. It explicitly models a specific sensor implementation where buffering occurs in the computer prior to the sensor. As a result, the measurement of the arrival rate becomes state dependent under high load. This property turns out to have some undesirable effects on the controlled system. The model was capable of predicting the behavior of the testbed quite well. Based on the presented model, analysis shows that feed-forward controllers suggested in the literature can lead to instability under certain circumstances at high load. This has not been reported earlier, but is in this doctoral thesis demonstrated by both simulations and experiments. The analysis explains why and when the instability arises. In the attempt to predict future response-times this thesis also presents a feedback based prediction scheme. Comparisons between earlier predictions to the real response-times are used to correct a model based response time prediction. The prediction scheme is applied to a controller to compensate for disturbances before the effect propagates to the response time. The method improves the transient response in the case of sudden changes in the arrival rate of requests. This doctoral thesis also presents work on a control solution for reserving CPU capacity for a given process or a given group of processes on a computer system. The method uses only existing operating-system infrastructure, and achieves the desired CPU capacity in a soft real-time manner.