Accelerating Particle-Mesh Algorithms with FPGAs and OmpSs@OpenCL

TL;DR

This paper explores FPGA-based acceleration of particle-mesh algorithms in plasma simulations, demonstrating promising performance and power efficiency improvements over GPU and CPU implementations.

Contribution

It introduces an FPGA hardware design for particle-mesh algorithms, leveraging deep pipelining and data parallelism, and compares its performance with GPU and CPU solutions.

Findings

FPGA implementation achieves higher power efficiency.

Hardware pipeline effectively handles dependencies during deposition.

Performance surpasses GPU and CPU counterparts in specific metrics.

Abstract

Due to its flexible architecture, FPGAs support unique, deep hardware pipeline implementations for accelerating HPC applications. However, these devices are quite new in the HPC space, and thus, have been scarcely explored outside some specific scientific domain, such as machine learning or biological sequence alignment. The objective of this thesis is to characterize the FPGA-based solution for accelerating particle-mesh algorithms, in which the force applied to each particle is computed based on the fields deposited in a finite mesh (or grid). Our starting point is a 2D kinetic PIC plasma simulator called ZPIC that has the same core algorithm and functionalities as OSIRIS. To create an efficient hardware design, the program keeps the particles strictly sorted by tiles (a group of cells) and uses the local memory as an explicitly managed cache. We also create multiple copies of the local current buffer to solve dependencies during the deposition phase. The resulting pipeline was replicated multiple times to explore data parallelism and increase its throughput. We then compare our hardware solution against similar implementations on GPU and multicore CPUs, showing promising results in term of power efficiency and performance. Keywords: FPGA, OpenCL, Kinetic Plasma Simulation.