Accelerating Scalability Analysis in the PaScal Analyzer with the Paramount Iteration Approach
Scalability Analysis, Performance Estimation, Paramount Iteration, Parallel Applications, PaScal Suite.
Scalability analysis can be challenging, as it often demands time and computational resources, given that it requires the complete execution of applications in different hardware configurations and with different problem sizes. To address this difficulty, this work presents a new feature in the PaScal Analyzer tool, which helps accelerate the collection of information needed for this analysis. The approach is based on the Paramount Iteration technique, which involves performing a partial execution of the parallel code to estimate the full performance. The method collects the time of the initial iterations, calculates the median of these values to predict the cost of the subsequent ones, and thus, terminates the application's execution prematurely. In initial results, using a matrix multiplication algorithm, the approach showed great potential, reducing analysis time by up to 89% with a low estimation error. Furthermore, the visual patterns of efficiency and scalability in the results were similar to those obtained with the application's full execution. The main advantage of this solution is that it offers flexibility to the user, making scalability analysis more practical, accessible, and faster.