Predição de desempenho de aplicações CUDA utilizando aprendizado de máquina e características de pré-execução
With the evolution of Graphics Processing Units (GPUs), parallel computing applications are becoming increasingly complex. Predicting the performance of these applications helps developers optimize their scheduling algorithms for workload distribution. In this work, machine learning models were deve...
| Autor principal: | SIQUEIRA, Luan Ribeiro |
|---|---|
| Grau: | Artigo |
| Publicado em: |
2024
|
| Assuntos: | |
| Acesso em linha: |
https://bdm.ufpa.br/jspui/handle/prefix/7423 |
| Resumo: |
|---|
|
With the evolution of Graphics Processing Units (GPUs), parallel computing applications are becoming increasingly complex. Predicting the performance of these applications helps developers optimize their scheduling algorithms for workload distribution. In this work, machine learning models were developed and evaluated to predict the performance of CUDA applications using pre-execution features. The Ridge Regression, Random Forest, and Decision Tree models were compared across nine CUDA applications using the MAPE metric. The results show that Decision Tree achieved the lowest MAPE values, while Random Forest demonstrated consistent performance. Ridge Regression had variable performance due to its limitation in handling multicollinearity. The study emphasizes the importance of considering the specific characteristics of the application and GPU when making performance predictions. |