Estimação adaptativa baseada em acurácia para tempo de execução de aplicações paralelas

Abstract

Scalability analysis allows the understanding of how efficiently a parallel application uses computing resources. This understanding provides insight into the behavior of the program, enabling optimizations, reducing resource consumption, and guiding development decisions for high-performance applications. The analysis process is very time-consuming because measuring application execution in several configurations is required. Estimating these execution times accurately is challenging due to the complex interactions between software and hardware, which can lead to problems such as load imbalance and resource contention. To deal with this problem, this paper proposes an adaptive approach using a multilayer perceptron neural network to predict the execution time of parallel algorithms iteratively by sampling different configurations of cores and input sizes as needed to ensure accuracy. Employing the approach on three different parallel algorithms (Matrix multiplication, Raytrace, and Breadth-first search), combined with three different pre-processing methods, validated the approach. The reduction in the time to perform the analysis was of up to 88.2%, with a maximum absolute error of 7%. Integrated with the Pascal Analyzer profiler, the model provides fast and accurate predictions, optimizing analysis time for researchers and developers. By accelerating scalability analysis and speedup calculation, the system makes the process of evaluating the performance of parallel applications more efficient, contributing to significant advances in high-performance computing.