Um estudo da extensão de paralelismo dinâmico fornecido por CUDA em aplicações recursivas

Abstract

The demand for application performance in the domains of graphical computation, machine learning, high-performance computing, etc. has been increasing over the years. To meet this demand, enterprises of such domains frequently tend to upgrade their computational resources, such as memory, processor, number of clusters, etc. However, this incremental upgrade may lead to a negative impact on costs related to energy. Therefore, one of the alternatives to increase the performance of applications without an increase in energy consumption of high-performance systems is the exploration of parallelism on the thread level. This level of parallelism can be explored in both homogeneous and heterogeneous architectures. Through the use of programming models that tend to make the programmer’s task simpler. One example of such a model is CUDA- (Compute Unified Device Architecture), a parallel programming model for heterogeneous environments composed of CPUs -(Central Processing Units) and GPUs -(Graphics Processing Units). In that sense, CUDA allows for the execution of parts or short segments of an application on the GPU. Through the use and creation of functions called kernels. By default, kernel calls need to be made on the CPU, and as such, there exists a cost associated with the communication between the CPU and GPU. In that sense, possible performance gains of recursive algorithms are limited. Given that, they require multiple kernel calls during their execution due to recursion. Besides that limitation, there also exists the necessity to use specialized kernels, to address the recursions of sub-problems of the graph. To solve this limitation, CUDA offers the programmer the extension dynamic parallelism, which allows kernel calls to be made directly on the GPU. This way, avoiding or minimizing the aforementioned cost. In this sense, this course conclusion work aims to evaluate the performance and energy consumption of the dynamic parallelism extension of CUDA in recursive algorithms.