High performance computing aspects of a dimension independent semi-Lagrangian discontinuous Galerkin code

TL;DR

This paper presents a parallel semi-Lagrangian discontinuous Galerkin method implementation that demonstrates high scalability and efficiency on large distributed memory systems, suitable for solving hyperbolic PDEs in multiple dimensions.

Contribution

It introduces a dimension-independent C++ framework for semi-Lagrangian discontinuous Galerkin methods with scalable parallel performance on clusters.

Findings

Achieved over 0.89 parallel efficiency up to 8192 cores in weak scaling.

Demonstrated good strong scaling up to 1024 cores.

Validated the framework with a Vlasov--Poisson solver.

Abstract

The recently developed semi-Lagrangian discontinuous Galerkin approach is used to discretize hyperbolic partial differential equations (usually first order equations). Since these methods are conservative, local in space, and able to limit numerical diffusion, they are considered a promising alternative to more traditional semi-Lagrangian schemes (which are usually based on polynomial or spline interpolation). In this paper, we consider a parallel implementation of a semi-Lagrangian discontinuous Galerkin method for distributed memory systems (so-called clusters). Both strong and weak scaling studies are performed on the Vienna Scientific Cluster 2 (VSC-2). In the case of weak scaling, up to 8192 cores, we observe a parallel efficiency above 0.89 for both two and four dimensional problems. Strong scaling results show good scalability to at least 1024 cores (we consider problems that can be run on a single processor in reasonable time). In addition, we study the scaling of a two dimensional Vlasov--Poisson solver that is implemented using the framework provided. All of the simulation are conducted in the context of worst case communication overhead; i.e., in a setting where the CFL number increases linearly with the problem size. The framework introduced in this paper facilitates a dimension independent implementation (based on C++ templates) of scientific codes using both an MPI and a hybrid approach to parallelization. We describe the essential ingredients of our implementation.