A dynamic approach for estimating service performance in the cloud

Abstract

Accurately estimating the service performance under a given resource configuration is of great importance to the resource provision for services in cloud platforms. To achieve this, it is necessary to build service performance models, the accuracy of which, however, is usually significantly influenced by the scale of training data. In this paper, combining collaborative filtering recommendation (CFR) and artificial neural network (ANN), we present a dynamic service performance modeling approach, called CADM, to improve the accuracy of estimation. In CADM, both performance models based on CFR and ANN are trained at service deployment time and runtime, and the one with lower mean absolute error is chosen to estimate the performance. Moreover, a merit-based threshold is introduced to reduce training costs. The experimental results illustrate that CADM has higher accuracy on different scales of training data, and the merit-based threshold has a significant impact on the estimation accuracy as well as the modeling efficiency.