Feedback control of internet applications involving the tracking of dynamic data

Abstract

Control of computing systems is different in many ways from the control of conventional systems. The quality of service problem in internet applications, that is, reducing the network overhead without affecting the quality, has been studied in detail. Network overhead can be measured in terms of the number of refreshes of the dynamic data. Use of randomly generated input vectors allows early use of feedback control, thereby lowering the network overhead. Identification using moving time windows helps obtain locally accurate models, thereby increasing the efficacy of the control scheme. Online tuning of the maximum permissible change in the input, namely maximum time till next refresh, has been shown to improve the control performance. Because this approach is not controller specific, it would be of use to all controllers. Two different ways of formulating the linear quadratic regulator objective function have been studied. Using the reciprocal of TTR (time to refresh) as the control input makes the control weight sensitive over a larger range, compared to using TTR itself as the input.