Enhanced Levenberg Marquardt Algorithm for Workload Prediction of Cloud Virtual Machine

Abstract

Cloud computing is the fast-growing technology which had been using widely for distrib uting and acquiring the required resources through internet. Due to incoming workload to cloud have nonlinear and inconsistent characteristics, predicting future workload remains a critical task. This causes, attempting to reduce numbers of running server at cloud data center also the key challenge in providing cloud services. One of the possible methods to overcome the issues is having prediction model which can increase accuracy level in dynamic workload environment. Purpose of future workload prediction is to provide reduction time window for deployment of cloud physical servers and virtual machine creation and their allocation. Neural network supervised machine learning model used for learning and predicting fu ture workload. Levenberg Marquardt algorithm is the combination of the steepest descent with low convergence and Gauss newton method with opposite characteristics. The major drawbacks of Levenberg Marquardt algorithm, it assumed independent of its initial point without proper ini tial weight initialization. Lack of method lead the model to many problems like, dead neurons, gradient disappearance, accuracy, and convergence problem. To increase the workload predic tion of virtual machine accuracy level, this paper designed proper weight initialization tech niques for ‘LM’ and implement in MATLAB R2016a. To show the performance of proposed sys tem, the result evaluated against ‘RNN’ algorithm where historical cloud virtual machine log data used for training in which CPU and Memory intensive workload used for performance met rics. From obtained result, we concluded the method show better accuracy than existing method in dynamic incoming workload to cloud datacenter.