Virtual machine load prediction model for agricultural cloud video platform based on semi-supervised partial least squares

Abstract: In order to optimize the infrastructure resource of agricultural cloud video platform efficiently, the virtual machine (VM) placement algorithms need to know the current and future efficiency of VM resource as accurately as possible for potential actions, such as service deployment, VM deployment, migration or cancellation. However, the data available for analysis are limited, as the samples used for prediction are usually very small. In the study, a sliding window model that considering time factor was designed to learn from small set of data. More importantly, the existing prediction algorithms still have much room to reduce the error. So a sliding window based mathematical method was provided to calculate the aforementioned forecasts, which was combined with PLS and semi-supervised learning (semi-supervised partial least squares, SS-PLS). The feasibility and advantages were analyzed in VM load forecasting based on SS-PLS method. Compared with auto regression moving average (ARMA), experimental results showed that the sliding window based model combined with SS-PLS made noticeable improvements to the forecasting accuracies, with root mean square error (RMSE) improved 5.47% to 1.777 86, mean absolute error (MAE) improved 6.37% to 1.331 2, and mean absolute percentage error (MAPE) improved 6.12% to 0.238 36, respectively. The results demonstrated that the proposed algorithm based on semi-supervised partial least squares is accurate in forecasting virtual machine load is effective in terms of the forecast accuracy.