Uma interface para escalonamento de algoritmos iterativos em C++

Abstract

Currently, parallel architectures are present in almost all kind of devices, justifying a search for tools and techniques that facilitate the development of parallel applications without jeopardizing performance. The current tools for parallel programming understand this need and focus their efforts on scheduling mechanisms that take advantage of common parallel patterns, present in the development of parallel applications. Iterative algorithms make up a large portion of computational problems, solutions that optimize the execution of these algorithms are related to the way the tasks are divided and scheduled. However, these tools do not offer the flexibility to change the scheduling policy, forcing the developer to solve problems according to the chosen tools and their way of scheduling, requiring the developer a knowledge of details about the implementation of tools to extract maximum performance. This approach offers a parallel object-oriented interface that allows to switch scheduling policies, allowing the scheduling strategy to be adapted according to the use of the algorithm. The parallel pattern map was used for a case study of performance validation and performance, comparing the results with other tools available for C ++, such as TBB and OpenMP.