基于CPU-GPU融合架构下的MapReduce编程模型

ata analysis and mining bring big challenges to traditional architecture of compute with increse of amount of data. With the continuous development of multi-core and parallel computing technology, people begin to pay close attention to GPU in the field of general-purpose computing, and many researches about GPU as co-processer to accelerate compute are put forward. In recent years, a fusion of GPU and CPU heterogeneous architecture obtained the widespread attention of the market because of its low power consumption and high performance computing ability . The parallel programming on a heterogeneous system has been confined by the complexity of the internal architecture of GPU. A distributed programming model - Mapreduce, provides a simple programming interface for programmer in big data computing and analysis. To simplify complexity of parallel programming on heterogeneous,we focus on using Mapreduce programming model on the new fusion CPU-GPU architecture in this paper. In order to make Mapreduce programming model adapt to the GPU features, we designed hash tables to manage intermediate key/value pairs without locking mechanism, and implemented a dynamic task scheduling method to fully utilize system computing period . We evaluated excuting time under four different workloads. The speedup compared with Mars and MapCG respectively outperform 1.4 to 3.6 and 1.3 to 26 times.