A robust QoS forecasting technique for a dynamic, distributed real-time testbed

Dynamic, distributed, real-time control systems must control changing environments in a timely manner despite the fact that the system's load and timing vary in a way that is not characterizable by time-invariant statistical distributions. A quality of service (QoS) manager has been implemented that forecasts timing constraint violations in such systems and corrects them before they occur. The majority of forecasting techniques rely on moving averaging to extrapolate the future values, therefore the existence of outliers frequently impose disastrous effects on the accuracy of prediction. Most existing forecasting methods in literature use thresholding steps to empirically eliminate outliers, whose success heavily depends on the prior knowledge in choosing the initial fit and threshold values. In this paper, we propose a robust algorithm to automatically reject outliers and thus achieve accurate forecasting of host load and path latency. Our algorithm involves minimizing the integral of the squared error (ISE or L2E) between a Gaussian model of the residual and its true density function. The residual here refers to the difference between the path latencies and the trend line. We present the implementation results using L2E as well as other two widely used forecasting methods: least-squares linear regression and Box-Jenkins AR(2) forecasting, with DynBench dynamic, distributed real-time benchmark being employed as the testbed. We experimentally show that our L2 E-based scheme yields higher forecasting accuracy over the other two approaches