One machine, one minute, three billion tetrahedra

TL;DR

This paper introduces a highly scalable parallel algorithm for 3D Delaunay triangulation, achieving over 55 million tetrahedra per second on multiple processors, enabling large-scale mesh generation.

Contribution

It presents a new efficient serial implementation and a multi-threaded parallelization scheme for 3D Delaunay triangulation, significantly improving performance and scalability.

Findings

Triangulated 3 billion tetrahedra in 53 seconds

Achieved over 55 million tetrahedra per second

Validated on three different processor architectures

Abstract

This paper presents a new scalable parallelization scheme to generate the 3D Delaunay triangulation of a given set of points. Our first contribution is an efficient serial implementation of the incremental Delaunay insertion algorithm. A simple dedicated data structure, an efficient sorting of the points and the optimization of the insertion algorithm have permitted to accelerate reference implementations by a factor three. Our second contribution is a multi-threaded version of the Delaunay kernel that is able to concurrently insert vertices. Moore curve coordinates are used to partition the point set, avoiding heavy synchronization overheads. Conflicts are managed by modifying the partitions with a simple rescaling of the space-filling curve. The performances of our implementation have been measured on three different processors, an Intel core-i7, an Intel Xeon Phi and an AMD EPYC, on which we have been able to compute 3 billion tetrahedra in 53 seconds. This corresponds to a generation rate of over 55 million tetrahedra per second. We finally show how this very efficient parallel Delaunay triangulation can be integrated in a Delaunay refinement mesh generator which takes as input the triangulated surface boundary of the volume to mesh.