Performance Evaluation of Cloud Centers with High Degree of Virtualization to provide MapReduce as Service

摘要

MapReduce has emerged as a paradigm where massive amountsof data are parallel processed with the help of clusters. Hadoop, asan open source implementation, has been used in a variety ofapplications such as social networking, video and image processing,log analysis and search indexing etc. For enterprises, providingMapReduce as a service in the cloud becomes an attractive model.Cloud Systems providing MapReduce as a service will allow users toaccess large number of machines in a cost-effectively mannerwithout creating the infrastructures of their own. MovingMapReduce to the virtualized environment will incur newchallenges, because the computation model is strongly bound todata, its storage, and location which make its behavior rather abatch processing. We consider cloud centers where tasks arrive inbatches or groups of random size and task service times are assumedto follow an exponential distribution. This paper also examines thecases where the arrival group size has a geometric distribution or adeterministic distribution. We examine a new analytical model forevaluation of performance of such large scale systems and computethe performance benchmarks such as mean waiting time in thequeue, mean request response time, mean system length and themean number of busy servers in the system