GeMiBi: A general multiple sources information Bayesian fusion for performance evaluation and an application to HPC cluster

Efficient and accurate performance evaluation is a challenge for many application areas. Information fusion is a widely used technology for this issue. Most existing information fusion methods have the requirement of taking a large sample into consideration. However, only small-scale experiments can be carried out for performance evaluation due to relatively severe resource constraints. To address this challenge, we delve into multiple sources information fusion method based on Bayesian inference for small samples case. In this paper, we propose GeMiBi: a general multiple sources information Bayesian inference method based on the minimum Jensen–Shannon Divergence (JSD). We exploit JSD to measure the similarity of different prior information and formulate a multiple constraints optimization problem to model the relation between different prior information and small samples observation data. In order to eliminate the massive numerical calculation when using the complex fused prior, we propose a novel and general information Bayesian inference method based on minimum JSD weights. Extensive experiments based on high performance cluster disk data are carried out to demonstrate the efficacy and effectiveness of the proposed method. Results show that the mean error of our method is 0.56% in the illustrating application, and it is greatly reduced compared with previous methods.

► The first one to study multiple sources information fusion based on small samples.
► A novel information Bayesian inference method based on minimum JSD to fuse the multiple sources is proposed.
► Experiments based on high performance cluster disk data are carried out to demonstrate the performance of our method.
► Results show that the mean error of our method is greatly reduced compared with previous methods.