Oral Presentation

Astaroth: Accelerating high-order finite-difference hydrodynamics with GPUs

Author(s): Miikka Väisälä (ASIAA, University of Helsinki); Johannes Pekkilä (Aalto University); Maarit Käpylä (MPS, Aalto University); Petri Käpylä (AIP, Aalto University, MPS); Omer Anjum (Aalto University, Nokia Solutions and Networks)

Presenter: Miikka Väisälä (Academia Sinica, Institute of Astronomy and Astrophysics)

We have have developed an experimental (magneto)hydrodynamics code Astaroth to accelerate 6th-order finite difference method, equivalent to the one used in the Pencil Code, with CUDA. I will discuss our 19- and 55-point implementations for Astaroth, focusing on challenges related to the high memory requirements which will lead to read-write latency issues, and what any astrophysical GPU code developer could learn from those challenges. A high-order method requires efficient uses of GPU memory types to provide substantial acceleration, particularly recycling values in shared memory and registers. While Astaroth was initial a stand-alone code it has been subsequently been integrated into Pencil Code, where it is available for public (https://github.com/pencil-code). Currently, Astaroth 2 is in development, which will provide greater flexibility and efficiency for scientific use.