Efficient time-energy execution of data-parallel applications on heterogeneous systems with GPU

摘要

This thesis proposes an approach for efficient time-energy execution of batch data-parallel applications on intra-node and intra-chip heterogeneous systems with GPU. This approach consists of three parts, namely, (i) techniques for efficient MapReduce data-parallel processing on heterogeneous systems with GPU, (ii) measurements for in-depth analysis of MapReduce data-parallel applications on intra-node and intra-chip heterogeneous systems and (iii) models to determine the execution time and energy usage of scale-out workloads and clusters. The techniques represented by lazy processing and dynamic mapping are implemented in MoSS using Hadoop and CUDA. MoSS is up to three times faster and up to 80% more energy-efficient compared to Hadoop. Moreover, both measurements and models show that heterogeneous systems with GPU are more time-energy-efficient that homogeneous systems, and that wimpy heterogeneous systems could potentially replace traditional brawny systems to achieve similar time performance while saving up to 90% of the energy.