Efficient Resource Prediction and Scheduling Approach for Scientific Applications in Cloud Environment

Abstract

Scientific Computing uses the state-of-the-art of high performance computing capabilities to solve the complex problems in various scientific domains such as weather forecasting, earthquake, sub-atomic particle behavior, turbulent flows and manufacturing processes etc. As the demand of resource requirements for solving the scientific problems is dy namic, so there is a need for a flexible platform which can handle the above-mentioned challenges in scientific applications concerning data storage and computation. Cloud computing provides a dynamic environment for deploying scientific applications by offering services such as infrastructure, platform and software. Various other features such as on-demand service, resource pooling, pay-as-per-use, elasticity, etc has attracted the scientists to deploy scientific applications on cloud. For effective utilization of virtualized resources in cloud, there is a need for efficient prediction based scheduling of tasks inorder to maximize performance and minimize execution time. Therefore, it is essential to first predict the resource requirements for scientific applications and then schedule them appropriately to meet the Quality of Service (QoS) requirements of the scientific users by taking SLA violations into consideration. To achieve the set objectives, an extensive literature survey of existing scientific applica tions has been done. Furthermore, state-of-the-art prediction techniques and scheduling approaches have been surveyed. From the literature, it can be inferred that prediction based scheduling is a challenging issue which needs to be handled carefully. To address these problems, firstly a Regressive Ensemble Approach for Predicting (REAP) resource usage has been proposed and based on the predicted set of resources a scheduling ap proach (RPS) has been devised. The results clearly state that the proposed ensemble approach (REAP) is better than the existing ones as it improves the overall accuracy, reduces the prediction time and yields the lowest error rate. Moreover, when compared with the existing scheduling heuristics of DataAware, FCFS, MaxMin, MinMin and MCT, the proposed approach schedules the tasks in the least execution time and maintains the cost and SLA violation rate at minimal levels.