Task-parallelism in SWIFT for heterogeneous compute architectures

TL;DR

This paper presents GPU acceleration techniques for the SWIFT hydrodynamics solver, achieving significant speedups and energy efficiency improvements through task-parallelism on heterogeneous architectures.

Contribution

It introduces novel algorithms enabling SWIFT to leverage task-parallelism on CPUs and GPUs simultaneously, optimizing performance and reducing communication bottlenecks.

Findings

GPU acceleration yields up to 3.5x speedup for offloaded computations.

Overall simulation speed increases by 1.8x on a superchip.

GPU acceleration improves energy efficiency by 29%.

Abstract

This paper highlights first steps towards enabling graphics processing unit (GPU) acceleration of the task-parallel smoothed particle hydrodynamics (SPH) solver SWIFT. Novel combinations of algorithms are presented, enabling SWIFT to function as a truly heterogeneous software leveraging task-parallelism on CPUs for memory-bound computations concurrently with GPUs for compute-bound computations while minimising the effects of CPU-GPU communication latency. The proposed algorithms are validated in extensive testing. The GPU acceleration methodology is shown to deliver up to 3.5 and 7.5 speedups for the offloaded computations when including and excluding the time required to prepare and post-process data transfers on the CPU side, respectively. The overall performance of the GPU-accelerated hydrodynamic solver for a full simulation on a single Grace-Hopper superchip is 1.8 times faster compared to the superchips fully parallelised CPU capabilities. This constitutes an improvement from 8 million particle updates/s for the full CPU-only baseline (115,000 updates per CPU core) to 15 million updates/s for the GPU-accelerated SPH solver. Moreover, it displays near-perfect strong scaling on 4 Grace-Hopper nodes. The GPU-acceleration is also demonstrated to give a 29 percent improvement in energy efficiency in comparison to CPU-only baselines. Finally, inter-influential bottlenecks in the prototype solver presented in this work are identified: A significant amount of time (up to 80 percent) of a GPU-offloading cycle is spent on preparing and post-processing particle data on the CPU for the transfer to and from the GPU, respectively. Approaches are suggested to minimise their effects and maximise the solver's performance in our future work.