Parallel prefix operations on heterogeneous platforms

Abstract

Craphics Processing Units (CPUs) have shown remarkable advantages in computing performance and energy efficiency, representing oue of the most promising trends fúr the near-fnture of high perfonnance computing. However, these devices also bring sorne programming complexities, and many efforts are required tú provide portability between different generations. Additionally, parallel prefix algorithms are a 8et of regular and highly-used parallel algorithms, whose efficiency is crutial in roany computer sCience applications. Although GPUs can accelerate the computation of such algorithms, they can also be a limitation when they do not match correctly to the CPU architecture or do not exploit the CPU parallelism properly. This dissertation presents two different perspectives. Gn the Oile hand, new parallel prefix algorithms have been algorithmicany designed for any paranel progrannning paradigm. On the other hand, a general tuning CPU methodology is proposed to provide an easy and portable mechanism tú efficiently implement paranel prefix algorithms on any CUDA CPU architecture, rather than focusing on a particular algorithm or a CPU mode!. To accomplish this goal, the methodology identifies the GPU parameters which influence on the performance and, following a set oí performance premises, obtains the cOllvillient values oí these parameters depending on the algorithm, the problem size and the CPU architecture. Additionally, the provided CPU functions are composed of modular and reusable CUDA blocks of code, which allow the easy implementation of any paranel prefix algorithm. Depending on the size of the dataset, three different approaches are proposed. The first two approaches solve small and medium-large datasets on a single GPU; whereas the third approach deals with extremely large datasets on a Multiple-CPU environment. OUT proposals provide very competitive performance, outperforming the stateof- the-art for many parallel prefix operatiOllS, such as the sean primitive, sorting and solving tridiagonal systems.