Key points are not available for this paper at this time.
The Sparse Matrix-Vector Multiplication (SpMV) serves as a critical computational kernel within sparse linear system solvers, playing a vital role in diverse scientific, engineering, and application domains such as industrial engineering, robotics, finance, networking, transportation, medicine, and weather forecasting. This study introduces a classification-based predictive model for shared memory, aiming to optimize scheduling policies in terms of execution time. Utilizing a dataset comprised of over 1000 matrices from 25 application domains, stored in block CSR format, the model is trained on 67% of the data using decision tree, random forest, gradient boosting, and artificial neural network algorithms. Results indicate that random forest exhibits superior accuracy with 20%, precision 4%, and recall 23% compared to other algorithms. Moreover, the runtime scheduling policy consistently demonstrates the best execution time across all matrices and features, outperforming static, dynamic, and guided policies. The study provides valuable insights into optimizing SpMV performance under various conditions, enhancing efficiency in solving sparse linear equation systems.
Ullah et al. (Fri,) studied this question.