Resource consumption prediction using neuro-fuzzy modeling

Abstract

The accurate prediction of resource consumption is important when it comes to optimally scheduling jobs in heterogeneous computer systems, e.g., cloud and grid computing infrastructures. Accordingly, different methods have been proposed to estimate the computer resource consumption of applications executed on such systems. In this paper, we use neuro-fuzzy modeling to predict the resource consumption of two bioinformatics applications, RAxML and BLAST. We experiment with different numbers and shapes of the membership functions to obtain, from a broad test set, the best initial configuration, which is tuned using neuro-adaptive learning methods. The results obtained by the neuro-fuzzy models are compared with those of five differently configured machine-learning models using the Root Relative Squared Error of a ten-fold cross validation of each model. This comparison indicates that neuro-fuzzy modeling can be used to estimate computer resource consumption and can provide more accurate or competitively accurate predictions of execution-time consumption.