An Experimental Analysis of Limitations of MapReduce for Iterative Algorithms on Spark

Abstract

MapReduce is the most popular framework for distributed processing. Recently, the scalability of data mining and machine learning algorithms has significantly improved with help from MapReduce. However, MapReduce does not handle iterative algorithms very efficiently. The problem is that many data mining and machine learning algorithms are iterative by nature. In order to overcome the limitations of MapReduce, many advanced distributed systems have been developed, including HaLoop, iMapReduce, Twister, and Spark. In this paper, we identify and categorize the limitations of MapReduce in handling iterative algorithms, and then, experimentally investigate the consequences of these limitations by using the most flexible and stable distributed system, Spark. According to our experiment results, the network I/O overhead was the primary factor that affected system performance the most. The disk I/O overhead also affected system performance, but it was less significant than the network I/O overhead. For the synchronization overhead, it affected system performance only when the static data was not cached.