Unified interface flux evaluation in a general discontinuous Galerkin spectral element framework

Boyang Xia, David Moxey

Introduction

High-order numerical methods, such as Discontinuous Galerkin Spectral Element Methods (DGSEM), are the "Formula 1" of computational fluid dynamics (CFD). They provide exceptional precision for tracking complex flow structures like vortices and turbulence. However, implementing these methods efficiently is notoriously difficult, particularly when the underlying mesh is non-conforming. Researchers at King’s College London have introduced a unified approach within the Nektar++ framework to handle these interfaces, balancing mathematical rigor with high-performance computing.

The Challenge of Non-Conforming Interfaces

In DGSEM, hp-refinement is often used to either reduce mesh size (h) or increase polynomial order (p) in regions of complex physics. This frequently creates "non-conforming" interfaces where elements of different polynomial orders meet or where mesh points do not align.

This presents two major challenges:

  1. Mathematical Symmetry: Methods like the Symmetric Interior Penalty Method (SIPG) require a symmetric system matrix. Inaccurate flux calculations on mismatched boundaries break this symmetry, causing efficient solvers like the Conjugate Gradient (CG) method to fail.
  2. Computational Efficiency: Traditional solutions, such as the mortar element method, are computationally expensive due to the need for data projection and large matrix inversions.

A Unified Framework for Flux Evaluation

The proposed framework implements two primary strategies within the open-source Nektar++ library to address these challenges while maintaining high performance.

Matrix-Free Implementation and Workflow

To maximize throughput, the authors avoid storing large matrices, opting instead for optimized "operators" like BwdTrans, PhysDeriv, and IProduct. These operators leverage SIMD vectorization and cache blocking to minimize slow memory access.

The Discontinuous Galerkin computational pipeline, illustrating the flow from volume flux kernels to trace flux evaluation and MPI communication.

The workflow, as shown above, separates volume and trace operations. The "trace-conforming" bridge allows the code to use standard high-performance operators even when physical interfaces are non-conforming.

Shared Trace Space and Interpolation

The framework utilizes a Shared Trace Space, a more efficient version of the mortar method. By defining a shared space and "imprinting" its quadrature points onto adjacent elements, the authors ensure both sides of an interface see the same integration points, preserving symmetry without heavy overhead.

For even faster processing, a Point-to-Point Interpolation approach is used. This requires precise quadrature to maintain symmetry when directly interpolating data between mismatched points.

The gather and scatter operations used to interpolate data between non-conforming element interfaces.

As illustrated in the interpolation logic, data is gathered from elements, interpolated to a common space, and then scattered back. This "unified" logic handles various element types (hexahedra, prisms, tetrahedra) consistently.

Key Results and Performance

The framework was validated using manufactured solutions and standard benchmarks:

  • Convergence: Expected $N_P$ convergence rates were achieved across all element types (Hex, Prism, Tet).
  • Symmetry Validation: The optimized approach reduced symmetry error to machine epsilon ($10^{-18}$), enabling the use of fast CG solvers. Naive interpolation, by contrast, resulted in errors around 2.5%.
  • Throughput: On modern EPYC 9554 processors, the implementation reached over $5 \times 10^9$ Degrees of Freedom (DOFs) per second for hexahedral elements.
  • Efficiency Trade-off: Direct interpolation was approximately 30% slower than the conforming case but significantly faster than complex projection methods.

Conclusion and Future Work

While the current work focuses on polynomial (p) non-conformance, it lays the groundwork for handling geometric (h) non-conformance. Future efforts will focus on extending these unified operators to GPUs and implementing more aggressive kernel fusing to push performance closer to theoretical hardware limits.

Reference: Xia, B., & Moxey, D. (2025). Unified interface flux evaluation in a general discontinuous Galerkin spectral element framework. arXiv:2504.03573v2 [math.NA].

Summary updated

BibTeX
@article{xia2026unified,
  author = {Boyang Xia and David Moxey},
  title = {{Unified interface flux evaluation in a general discontinuous Galerkin spectral element framework}},
  journal = {SIAM J. Sci. Comput.},
  year = {2026},
  doi = {10.1137/25m174824x},
}
Figure