DescriptionGiven the large number of GPU architectures and the many different possibilities to execute an algorithm, the prediction of application execution times over GPUs is a great challenge and is essential for efficient JMSs. The available GPU performance modeling solutions are very sensitive to applications and platform changes,
Here a summary of two main works is shown. In the first work, we present the comparison of a developed BSP-based model to three different ML techniques, this comparison was done with 9 well-known matrix/vector applications. In this work, we wanted to perform a fair comparison, for this reason, we decided that ML process would had the same features that the BSP-based model.
In the second work, we have compared among ML techniques. Here, a two step of extraction features are done. First a correlation analysis and after hierarchical clustering analysis. In this second work, 10 irregular CUDA kernels were used.