Scalability of Gaussian Processes Using Asynchronous Tasks: A Comparison Between HPX and PETSc.

Gaussian processes are a widely used alternative to neural networks for non-linear system identification. The method requires computing the inversion of a large covariance matrix. In this work, we introduce our new task-based asynchronous implementation, focusing on its most popular solver, the Cholesky decomposition. Our implementation is based on HPX, utilizing its asynchronous many-task runtime system. We can therefore investigate its scaling on multi-core hardware and for GPU offloading. Furthermore, we compare our HPX implementation against a high-level reference implementation based on PETSc. We demonstrate that the HPX implementation’s performance is directly tied to the chosen tile size. Compared to the PETSc reference, our task-based implementation is faster in the entire node-level strong scaling experiment on EPYC ROME, showing better parallel efficiency.