Meparalel : um método para análise de implementação de algoritmo paralelo baseado em CUDA

Abstract

There are basically two approaches to attempt to improve performance of the algorithms: (i) the hardware-based and (ii) the software-based. The approaches based on software, that before were based on sequences algorithms, could not extract the hardware resources available. To solve this problem the parallel algorithms arose. Parallel algorithms tend to do their jobs more quickly due to their ability to distribute their workload by the available multi-core processors. In the search for the processing improvement the GPU started to be used in general purpose computing, and changed from a simple graphics processor to a parallel coprocessor capable of simultaneously performing thousands of operations. NVIDIA to popularize the GPU use in general purpose computing launched the CUDA which allows developers to parallelize their solutions more intuitively. But it is not an easy task to parallelize in order to improve resources utilization and reduce the processing time. Thus, as the literature offers no suitable mechanism, this paper proposes a method for analysis of parallel algorithms that can help the process of analysis and refactoring code built in CUDA programming platform and what can generate faster, more efficient algorithms in the consumption of hardware resources.