SAAP: A State-Aware Adaptive Prediction Strategy for CPU Load of Desktops

Most medium-scale above corporations usually construct their own private clouds, utilizing commodity servers to provide computing resources. With the increase in computing demand, more and more servers are required. On the other hand, in these corporations, there are hundreds or thousands of desktops (physical or virtual personal computers) that are running with low resource utilizations. With the availability of container technologies, such as Docker, it is currently feasible to utilize these potential computing resources. To do so, it is critical to predict the resource consumption of desktops accurately before scheduling jobs for them, in order to improve the execution of jobs. Although some approaches have been proposed to predict the resource utilization of servers, they can't be directly applied to desktops due to the dynamics of desktops. To address this problem, we propose SAAP, a State-Aware Adaptive Prediction strategy for CPU load of desktops. SAAP is capable of dynamically selecting appropriate prediction algorithms to predict the CPU load, adapting to the state of desktops. Besides, two patterns that can improve prediction accuracy are found. To evaluate the effectiveness of SAAP, extensive experiments are conducted. The experimental results demonstrate that SAAP behaves much better than the Box-Jenkins models (AR, MA, ARMA, ARIMA) in prediction accuracy.