Performance evaluation and relative predictive model of parallel file system

Abstract

In this paper, the performance evaluation and modeling of parallel file system based on Lustre file system is studied. After performing a survey on performance factors, a series of performance evaluations via experimental approaches and propose a performance relational model (PRModel). In the experimental and PRModel analysis, it is found that different performance factors have closed performance correlations. In order to mime the relational information, a novel relative performance predictive model (RPPModel) is proposed. This model can be used to predict the overhead over different performance factors. The model is validated through a series of experiments over a variety of performance factors. The experimental results show that the average relative errors results can be controlled within 17%~28%. This model is easy to use and can obtain better prediction accuracy. ©2011, Institute of Software, the Chinese Academy of Sciences. All rights reserved.