Locality Based Data Partitioning in MapReduce

Performance of MapReduce heavily depends on its data partitions for parallelism controlling, however, current state of art solutions are implemented using naive methods which are far from optimal. In this paper, we outline our solution, a locality based and skew aware partition technology, intend to address the data partition problems which encumber the job performance seriously. Our solution LBP(Locality Based Partitioning) clusters data blocks from a same node into a single partition, which need only one map task to process avoiding the spoil time for slot reallocation and multiple tasks initializing. Realizing the data skew problem, we enhance the LBP to LBP-SA(LBP Skew Aware) to partition the data file according their record and computation skews, so decrease the variety of tasks lifetime. Experiments results demonstrated that our solutions can improve the MapReduce processing performance remarkably than traditional Hadoop Implementation.