Abstract Views :269 |
PDF Views:0
Authors
Affiliations
1 School of Computing Science and Engineering, VIT University, Vellore – 632014, Tamil Nadu, IN
2 School of Electrical Engineering, VIT University, Vellore – 632014, Tamil Nadu, IN
Source
Indian Journal of Science and Technology, Vol 8, No 35 (2015), Pagination:
Abstract
Background/Objectives: To evaluate the prediction accuracy of Neural Network algorithms for host CPU load prediction and evaluate their performance compared to actual values. Methods/Statistical Analysis: The speed of execution of job at the scheduled host is directly proportional to its CPU load. Therefore, target node load prediction plays an important role in job scheduling decisions. It is learnt that Neural Networks are capable of predicting the future values based on the training given on the past data. We designed a multilayer neural network and trained with learning algorithms for the input patterns collected from the load traces and predicted the future load statistics. The Mean and Standard Deviation of the predicted values are computed and analyzed against the Mean and Standard Deviation of actual values for all the ANN algorithms. Findings: We analyzed the prediction accuracy of Back Propagation, Quick Propagation, Back Propagation with Momentum and Resilient Propagation algorithm for the load traces collected from variety of computers connected in a network. Existing reports shows that Back Propagation algorithm exhibits better prediction accuracy compared to statistical approaches like linear regression and polynomial regression. In this paper, we have shown that Resilient Propagation algorithm has better prediction accuracy compared to other ANN algorithms. Application/Improvements: Job scheduling and resource selection algorithms can employ neural network algorithms to predict the load for the sharable resources connected in the network for more accurate and faster scheduling/resource selection decision.
Keywords
CPU Load Prediction, Load Traces, Neural Network Algorithms, Training, Resource Selection
Full Text