Performance Analysis Using Petri Net Based MapReduce Model in Heterogeneous Clusters

Currently, big data and large-scale data processing techniques has become an important developing area. MapReduce is an enabling technology of cloud computing. Hadoop is one of the most popular MapReduce implementation, which is the target platform in this paper. When running a MapReduce job, programmers however cannot acquire the information about how to fine-tune the parameters of application. Moreover, programmers need much time on finding the most suitable parameters. This paper evaluates execution processes in MapReduce and form SPN-MR model with Stochastic Petri Net. In order to analyze the performance of SPN-MR, formulas of mean delay time in each time transition are defined. SPN-MR simulates the elapsed time of any MapReduce jobs with known input data sizes and then reduces time cost in performance tuning. SPN-MR carried out several actual test benchmarks. The results showed the average error rate is within 5 percent. Therefore, it can provide effective performance evaluation reports for MapReduce programmers.