Astrophysical code migration into Exascale Era

TL;DR

This paper discusses adapting astrophysical N-body simulation code for exascale computing platforms using ARM64 cores and FPGA accelerators, including code optimization, porting, and initial performance testing on ARM SoC hardware.

Contribution

It presents novel re-engineering and porting strategies for Hy-Nbody code to ARM64 and FPGA platforms targeting exascale readiness, with initial performance insights.

Findings

Hy-Nbody code was successfully optimized for ARM64 architecture.

Initial tests on ARM SoC show promising performance results.

Porting to FPGA enables performance-per-watt comparisons.

Abstract

The ExaNeSt and EuroExa H2020 EU-funded projects aim to design and develop an exascale ready computing platform prototype based on low-energy-consumption ARM64 cores and FPGA accelerators. We participate in the application-driven design of the hardware solutions and prototype validation. To carry on this work we are using, among others, Hy-Nbody, a state-of-the-art direct N-body code. Core algorithms of Hy-Nbody have been improved in such a way to increasingly fit them to the exascale target platform. Waiting for the ExaNest prototype release, we are performing tests and code tuning operations on an ARM64 SoC facility: a SLURM managed HPC cluster based on 64-bit ARMv8 Cortex-A72/Cortex-A53 core design and powered by a Mali-T864 embedded GPU. In parallel, we are porting a kernel of Hy-Nbody on FPGA aiming to test and compare the performance-per-watt of our algorithms on different platforms. In this paper we describe how we re-engineered the application and we show first results on ARM SoC.