Statistical and Artificial Neural Network Based Modeling of Parallel Job Scheduling Algorithms

Abstract

The prime applicability of parallel space-sharing job scheduling algorithms in PC-cluster is to schedule jobs and efficiently allocate cluster's processors to the jobs to achieve performance objective viz. minimized average turnaround time (ATT). This paper demonstrates the use of the two-phase strategy based on the Response Surface Methodology (RSM) approach of Design of Experiments (DOE) and Artificial Neural Networks (ANN) for modeling the performance of parallel space-sharing job scheduling algorithms particularly Largest Job First (LJF) algorithm. In the first phase DOE based statistical-mathematical techniques helps in identifying, ranking and modeling the significant independent scheduling process variables affecting the ATT based output values with minimal cost involved in terms of experimental runs, money and time. RSM based regression analysis helps to fit second-order quadratic empirical model equation for output metric ATT involving main and interaction effects terms of scheduling process variables. High values of coefficient of determination R2, adjusted R2 and insignificant lack of fit represent the goodness of fit of the model to accurately model the ATT values. In the second phase ANN model for ATT is developed using the experimental data passed from DOE phase to validate the RSM based model predictions. The two-phase modeling strategy tends to combines the advantages of RSM and ANN approaches.

Keywords: PC-cluster, Largest job first, Design of experiments, Response surface methodology, Average turnaround time, Artificial neural networks