High-performance low-dissipation algorithms for simulation of turbulent compressible flows
(English) Motivated by recent advances in computational technology aiming at exascale capabilities, which implies a need for applications capable of taking advantage of these new supercomputing archit ectures, this work will present two algorithms aimed at implementing an efficient and low-dissipati...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/693381 |
| Acceso en línea: | http://hdl.handle.net/10803/693381 https://dx.doi.org/10.5821/dissertation-2117-424242 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Aeronàutica i espai 531/534 |
| Sumario: | (English) Motivated by recent advances in computational technology aiming at exascale capabilities, which implies a need for applications capable of taking advantage of these new supercomputing archit ectures, this work will present two algorithms aimed at implementing an efficient and low-dissipation algorithm focused on LES and DNS of turbulent compressible flows. The basis for the algorithms is the Continuous Galerkin method applied to elements whose nodes and quadrature points are formed from the Gauss-Lobatto-Legendre roots, resulting in a Spectral Elements Method. Throughout this work, it will be evidenced that this formulation leads to highly efficient kernels for discretizing the convective and diffusive terms of the compressible N avier-Stokes equations, with the added benefit that the order of the scheme is coupled with the order of the shape function polynomials employed by the elements themselves, significantly sim plifying the process of increasing the order of the scheme. To achieve a stable Total Variational Diminishing algorithm, the \acrshort{sem} scheme will be paired with an Entropy Viscosity-based stabilization model and a suitable splitting of the nonl inear convective terms will be employed to eliminate aliasing issues that occur in the \acrshort{sem} formulation. This spatial discretization is then coupled with both an explicit and a semi-implicit scheme to account for the temporal nature of the flow equations. The explicit version of the algorithm i s expected to be simple and efficient per time step, but due to its \acrshort{cfl} condition limitation, the semi-implicit version is also proposed to allow for better overall performance in cases where the time-step becomes overly limited, such as in wall-bounded flows. Considering the focus on producing a \acrshort{cfd} application towards the exascale future, an important aspect of this work is that both algorithms are proposed with a full \acrshort{gpu} implementation in mind: the use of accelerators is expected to be a key aspect of future supercomputing architectures, and thus it is important to design these algorithms with this in mind. Examples detailing the performance of both algorithms will be presented both in the case of a single device and when distributed architectures using multiple devices are employed. |
|---|