A Model for Software Measurement Aiming to Guide Evaluations and Comparisons between Programming Tools to Implement GPU Applications

Abstract

Programming tools for GPUs are frameworks that offer resources to explore the massive parallelism power provided by these devices. Nowadays, they are being extensively applied for HPC purposes. Despite the existence of many tools that can be used to encode a GPU program, programming targeting GPUs is still seen as challenging, requiring the use of specialized frameworks to deal with the heteroge neous environment demanded by the GPU architecture. Also, there isn’t a commonly accepted definition of a standard framework for GPU programming. So, the process to select a tool in order to implement a GPU program is not simple, especially when this choice can impact the performance and programming effort required to implement it. Regarding that, this Thesis proposes a model to guide evaluations and comparisons between frameworks for GPUs. This model was designed based on the GQM method for software measurement and, because of that, was formulated using goals, ques tions and metrics to analyze three different aspects about the frameworks: program ming expressiveness, programming effort and performance. As a result, the model seeks to offer a perspective including the characteristics, strengths and weaknesses about programming tools. We have in mind that such perspective can support the pro cess of choosing a framework to develop a GPU program. Experiments guided by the proposed GQM model were conducted applying the NAS Parallel Benchmarks imple mented with CUDA, OpenACC and Python/Numba. The experimental results contem plated the three aspects defined in the model, showing the similarities and differences regarding the tested APIs. Further, these results were employed to compose the per spective considering each framework. Due that, we believe this study contributes to improving the available knowledge about these tools, as well as advancing research on programming tools for GPUs.