A Machine Learning-Based Approach for Predicting the Execution Time of CFD Applications on Cloud Computing Environment

Abstract

Calibrations and validations of Computational Fluid Dynamics (CFD) applications are significantly time-consuming. To reduce the execution time of the CFD applications, parallel-computing approach is often employed. In addition, high performance computing systems and cloud computing solutions are also appropriate tools to the CFD applications. One of the challenging problems is to schedule tasks on virtualized machines of the cloud-based high performance systems. Instead of employing an adaptive algorithm to cope with the uncertainty of the virtualized resources, in this study, we propose an idea to predict the execution time of Telemac-2D, which is a CFD application. The predicted execution time is very essential in all scheduling algorithms. The application is executed several times with different settings of model’s parameters and allocated resources to produce an experimental dataset. The dataset is then used to predict the execution time of the application by utilizing a machine learning-based approach. The predictive model consists of two steps that classify and predict the execution. The C4.5 algorithm is used to classify the execution ending status whereas Multi-layer Perceptron (MLP) and a mixture of MLPs (MiMLP) are used to predict the execution time. The experiments indicate that the predictive model is appropriate to predict the execution of the Telemac-2D application since the accuracy of the C4.5 algorithm is 100 % and R and MARE of MiMLP are 0.957 and 17.090, respectively.