Extension of Parallel Primitives and Their Applications to Large-Scale Data Processing

Data parallel primitives are highly optimized general-purpose algorithms designed only for GPUs and are used as building blocks to develop applications. However, existing data parallel primitives cannot handle data larger than the GPU memory size. In this paper, we propose an extension to existing data parallel primitives to efficiently handle data larger than the GPU memory size by cooperatively using both GPUs and CPUs. Moreover, we evaluate the impact of these primitives when applying them to large data processing applications, with respect to both performance and software development cost.