Abstract: The relationship model between resources and service performance is a key to the proper virtual resource allocation for services in cloud environment. However， the accuracy of these non-linear relationship models is usually significantly influenced by the scale of training data. Aiming at the shortcomings of related work， a dynamic service performance modeling method named CSDM， which combines collaborative filtering recommendation and support vector regression， was proposed. In CSDM， for better accuracy， both performance models were trained at service deployment time and runtime， and the one with lower MAE was selected to estimate the performance under given resource status. In addition， a merit-based threshold was introduced to reduce training costs of performance models. The experimental results showed that CSDM had higher accuracy on different scales of training data， and the merit-based threshold had a significant effect on the prediction accuracy as well as the modeling efficiency.

Key words: cloud service, performance model, resource status, CFR(collaborative filtering recommendation), SVR(support vector regression)