Utilização de técnicas de análise estática e dinâmica para otimização de aplicações de propósito geral em GPUs

Abstract

We have witnessed large advances in the use of devices designedspecifically for running parallel applications. Examples of this class of hardware are graphics processing cards, as those produced by NVIDIA.The use of GPUs to accelerate general purpose applications gave rise to a new line of development parallel applications, popularly known as GPGPU. Although recent, GPGPU applications achieved performance up to 100 times faster then traditional programs. However, programming in these environments presents several challenges, due to factors such as the complicated arrangement of memory, large flexibility when scheduling threads and the very nature of applications that need to be parallelized, which often is not fit into GPU\\\'s SIMD programming model. In order to mitigate these challenges, allowing non-specialist developers to take advantage of the large computing power of the GPGPU environment, we propose new techniques for performance evaluation and optimization in this environment.Our techniques are based on the combined use of static and dynamic program analysis. Static analysis techniques provide the developer a global vision the parallel application, but this view is often unnecessarily conservative. Moreover, dynamic analysis techniques provide a accurate view of the application under review, however this view is local and depends on factors such as input and execution environment. We hope that these two universes of program analysis can be combined so that developers can obtain more information than is possible using each strategy separately.