Cache-conscious Splitting of MapReduce Tasks and its Application to Stencil Computations

Modern cluster systems are typically composed by nodes with multiple processing units and memory hierarchies comprising multiple cache levels of various sizes. To leverage the full potential of these architectures it is necessary to explore concepts such as parallel programming and the layout of data onto the memory hierarchy. However, the inherent complexity of these concepts and the heterogeneity of the target architectures raises several challenges at application development and performance portability levels, respectively. In what concerns parallel programming, several model and frameworks are available, of which MapReduce [16] is one of the most popular. It was developed at Google [16] for the parallel and distributed processing of large amounts of data in large clusters of commodity machines. Although being very powerful tools, the reference MapReduce frameworks, such as Hadoop and Spark, do not leverage the characteristics of the underlying memory hierarchy. This shortcoming is particularly noticeable in computations that benefit from temporal locality, such as stencil computations. In this context, the goal of this thesis is to improve the performance of MapReduce computations that benefit from temporal locality. To that end we optimize the mapping of MapReduce computations in a machine’s cache memory hierarchy by applying cacheaware tiling techniques. We prototyped our solution on top of the framework Hadoop MapReduce, incorporating a cache-awareness in the splitting stage. To validate our solution and assess its benefits, we developed an API for expressing stencil computations on top the developed framework. The experimental results show that, for a typical stencil computation, our solution delivers an average speed-up of 1.77 while reaching a peek speed-up of 3.2. These findings allows us to conclude that cacheaware decomposition of MapReduce computations considerably boosts the execution of this class of MapReduce computations