Modeling and Performance Evaluation of MapReduce in Cloud Computing Systems Using Queueing Network Model

Abstract

MapReduce is a two -stage information processing technique and it is common concept for big data. Map and Reduce procedures are distributed among some processors within a cluster in the cloud. The performance modeling and analysis of MapReduce execution times has been a challenging task. Analytic performance models provide reasonably accurate job response time estimation with significantly lower cost compared with experimental experiments. Queueing theory is one the modeling and analysis tools of such systems since it enables efficient analysis of the performance, availability and some other key metrics of a data processing system. In this paper, an M/G/1/K performance model with first come first serviced (FCFS) discipline of MapReduce is proposed. More specifically, it will present a queueing model with two stages hypoexponential service time and finite queue. This model has a cloud server with two stages to investigate the performance of the MapReduce technique subject to heavy traffic conditions. The system is analyzed via discrete-event simulation (DES). Key numerical examples are presented for varying number of mappers, reducers and the mean arrival rates to assess their effect on the system mean response time, loss probability and mean queue length. The results are expected to be useful for predicting MapReduce under various workloads and operating conditions of big data processing.