Runtime prediction models for web-based system resources

Several activities of Web-based architectures are man-aged by algorithms that take runtime decisions on the ba-sis of continuous information about the state of the internal system resources. The problem is that in this extremely dy-namic context the observed data points are characterized by high variability, dispersion and noise at different time scales to the extent that existing models cannot guarantee accurate predictions at runtime. In this paper, we evaluate the predictability of the internal resource state and point out the necessity to filter the noise of raw data measures. We then verify that more accurate prediction models are re-quired which take into account the non stationary effects of the data sets, the time series trends and the runtime con-straints. To these purposes, we propose a new prediction model, called trend-aware regression. It is specifically de-signed to deal with on the fly and short-term forecast of time series which originate from filtered data points belonging to internal resources of Web system. The experiment evalua-tion for different workload scenarios shows that the pro-posed trend-aware regression model improves the predic-tion accuracy with respect to popular algorithms based on auto-regressive and linear models, while satisfying the com-putational constraints of runtime prediction.