Accurate memory resource prediction can achieve optimal performance for complex system simulation (CSS) using optimistic parallel execution in the cloud computing environment. However, because of the varying memory resource demands of CSS applications caused by the simulation entity scale and frequent optimistic synchronization, the existing approaches are unable to predict the memory resource required by a CSS application accurately, which cannot take full advantage of the elasticity and symmetry of cloud computing. In this paper, a probabilistic prediction approach based on ensemble learning, which regards the entity scale and frequent optimistic synchronization as the important features, is proposed. The approach using stacking strategy consists of a two-layer architecture. The first-layer architecture includes two kinds of base models, namely, back-propagation neural network (BPNN) and random forest (RF). The root mean squared error-based pruning algorithm is designed to choose the optimal subset of the base models. The second-layer is the Gaussian process regression (GPR) model, which is applied to quantify the uncertainty information in the probabilistic prediction for memory resources. A series of experiments are presented to prove that the proposed approach can achieve higher accuracy and performance compared to RF, BPNN, GPR, Bagging ensemble approach, and Regressive Ensemble Approach for Prediction.
Optimistic Synchronization-Based State-Space Reduction
