high performance computing on graphics processing units: hgpu.org

The high performance computing landscape is shifting from assemblies of homogeneous nodes towards heterogeneous systems, in which nodes consist of a combination of traditional out-oforder execution cores and accelerator devices. Accelerators, built around GPUs, many-core chips, or FPGAs, are used to offload compute-intensive tasks. These devices provide superior theoretical performance compared to traditional multi-core CPUs, but not every application fits into the programming model they impose, and exploiting their computing power remains a challenging task. This dissertation discusses the issues that arise when trying to efficiently use general purpose accelerators. As a contribution to aid in this task, we present a thorough survey of performance modeling techniques and tools for general purpose coprocessors. Then we use as case study the statistical technique Kernel Density Estimation (KDE). KDE is a memory bound application that poses several challenges for its adaptation to the accelerator-based model. We present a novel algorithm for the computation of KDE that reduces considerably its computational complexity, called S-KDE. Furthermore, we have carried out two parallel implementations of S-KDE, one for multi and many-core processors, and another one for accelerators. The latter has been implemented in OpenCL in order to make it portable across a wide range of devices. We have evaluated the performance of each implementation of S-KDE in a variety of architectures, trying to highlight the bottlenecks and the limits that the code reaches in each device. Finally, we present an application of our S-KDE algorithm in the field of climatology: a novel methodology for the evaluation of environmental models.