MAPSkew: Metaheuristic Approaches for Partitioning Skew in MapReduce

Abstract

MapReduce is a parallel computing model in which a large dataset is split into smaller parts and executed on multiple machines. Due to its simplicity, MapReduce has been widely used in various applications domains. MapReduce can significantly reduce the processing time of a large amount of data by dividing the dataset into smaller parts and processing them in parallel in multiple machines. However, when data are not uniformly distributed, we have the so called partitioning skew, where the allocation of tasks to machines becomes unbalanced, either by the distribution function splitting the dataset unevenly or because a part of the data is more complex and requires greater computational effort. To solve this problem, we propose an approach based on metaheuristics. For evaluating purposes, three metaheuristics were implemented: Simulated Annealing, Local Beam Search and Stochastic Beam Search. Our experimental evaluation, using a MapReduce implementation of the Bron-Kerbosch Clique Algorithm, shows that the proposed method can find good partitionings while better balancing data among machines.

Keywords

• MapReduce
• partitioning skew
• metaheuristics
• parallel computing
• high performance computing