On The Performance, Energy, And Power Of Data-Access Methods In Heterogeneous Computing Systems

Graphics processing units (GPUs) have delivered promising speedups in data-parallel applications. A discrete GPU resides on the PCIe interface and has traditionally required data to be moved from the host memory to the GPU memory via PCIe. In certain applications, the overhead of these data transfers between memory spaces can nullify any performance gains achieved from faster computation on the GPU. Recent advances allow GPUs to directly access data from the host memory across the PCIe bus, thereby alleviating the data-transfer bottlenecks.Another class of accelerators called accelerated processing units (APUs) mitigate data-transfer overhead by placing CPU and GPU cores on the same physical die. However, APUs in the current form provide several different data paths between the CPU and GPU, all of which can differently affect application performance. In this paper, we explore the effects of different available data paths on both GPUs and APUs in the context of a broader set of computation and communication patterns commonly referred to as dwarfs.