In recent years the increasing availability of High Performance Computing (HPC) resources strongly promoted the widespread of high fidelity simulations, such as the Large Eddy Simulation (LES), for industrial research and design. One of the most promising approaches to those kind of simulations is based on the discontinuous Galerkin (dG) discretization method. The contribution of the thesis towards this research area is three-fold. First, the work introduces an efficient hybrid MPI/OpenMP parallelisation paradigm to fruitfully exploit large HPC facilities. Second, it reports efficient, scalable and memory saving solution strategies for stiff dG discretisations. Third, it compares those solution strategies, for the first time using the same numerical framework, to hybridizable discontinuous Galerkin (HDG) methods, including a novel implementation of a p-multigrid preconditioning approach, on unsteady flow problems involving the solution of the NavierStokes equations. The improvements in computational efficiency have been evaluated on cases of growing complexity involving large eddy simulations of turbulent flows. First, the Rayleigh-Benard convection problem and the turbulent channel flow at moderately high Reynolds numbers is presented. The solution strategies proposed resulted up to five times faster than standard matrix-based methods while al- locating the 7% of the memory. A second family of test cases involve the LES simulation of a rounded leading edge flat plate under different levels of free-stream turbulence. Although the increased stiffness of the iteration matrix due to the use of curved and stretched elements, the solver resulted more than three times faster while allocating the 15% of the memory if compared to standard methods. Finally, the large eddy simulation of the Boeing Rudimentary Landing Gear at Re = 10^6 is reported. In all the cases, a remarkable agreement with experimental data as well as previous numerical simulations is documented.
Negli ultimi anni, la crescente disponibilit`a di risorse computazionali ha contribuito alla diffusione della fluidodinamica computazionale per la ricerca e per la progettazione industriale. Uno degli approcci pi promettenti si basa sul metodo agli elementi finiti discontinui di Galerkin (dG). Nell’ambito di queste metodologie, il contributo della tesi e' triplice. Innanzi- tutto, il lavoro introduce un algoritmo di parallelizzazione ibrida MPI/OpenMP per l’utilizzo efficiente di risorse di super calcolo. In secondo luogo, propone strategie di soluzione efficienti, scalabili e con limitata allocazione di memoria per la soluzione di problemi complessi. Infine, confronta le strategie di soluzione introdotte con nuove tecniche di discretizzazione dette “ibridizzabili”, su problemi riguardanti la soluzione delle equazioni di Navier–Stokes non stazionarie. L’efficienza computazionale e' stata valutata su casi di crescente complessita' riguardanti la simulazione della turbolenza. In primo luogo, e' stata considerata la convezione naturale di Rayleigh-Benard e il flusso turbolento in un canale a numeri di Reynolds moderatamente alti. Le strategie di soluzione proposte sono risultate fino a cinque volte piu` veloci rispetto ai metodi standard allocando solamente il 7% della memoria. In secondo luogo, e' stato analizzato il flusso attorno ad una piastra piana con bordo arrotondato sottoposta a diversi livelli di turbolenza in ingresso. Nonostante la maggiore complessità' dovuta all’uso di elementi curvi ed anisotropi, l’algoritmo proposto e' risultato oltre tre volte piu` veloce allocando il 15% della memoria rispetto ad un metodo standard. Concludendo, viene riportata la simulazione del “Boeing Rudimentary Landing Gear” a Re = 10^6. In tutti i casi i risultati ottenuti sono in ottimo accordo con i dati sperimentali e con precedenti simulazioni numeriche pubblicate in letteratura.
Towards Efficient and Scalable Discontinuous Galerkin Methods for Unsteady Flows
FRANCIOLINI, MATTEO
2019
Abstract
In recent years the increasing availability of High Performance Computing (HPC) resources strongly promoted the widespread of high fidelity simulations, such as the Large Eddy Simulation (LES), for industrial research and design. One of the most promising approaches to those kind of simulations is based on the discontinuous Galerkin (dG) discretization method. The contribution of the thesis towards this research area is three-fold. First, the work introduces an efficient hybrid MPI/OpenMP parallelisation paradigm to fruitfully exploit large HPC facilities. Second, it reports efficient, scalable and memory saving solution strategies for stiff dG discretisations. Third, it compares those solution strategies, for the first time using the same numerical framework, to hybridizable discontinuous Galerkin (HDG) methods, including a novel implementation of a p-multigrid preconditioning approach, on unsteady flow problems involving the solution of the NavierStokes equations. The improvements in computational efficiency have been evaluated on cases of growing complexity involving large eddy simulations of turbulent flows. First, the Rayleigh-Benard convection problem and the turbulent channel flow at moderately high Reynolds numbers is presented. The solution strategies proposed resulted up to five times faster than standard matrix-based methods while al- locating the 7% of the memory. A second family of test cases involve the LES simulation of a rounded leading edge flat plate under different levels of free-stream turbulence. Although the increased stiffness of the iteration matrix due to the use of curved and stretched elements, the solver resulted more than three times faster while allocating the 15% of the memory if compared to standard methods. Finally, the large eddy simulation of the Boeing Rudimentary Landing Gear at Re = 10^6 is reported. In all the cases, a remarkable agreement with experimental data as well as previous numerical simulations is documented.File | Dimensione | Formato | |
---|---|---|---|
Tesi_Franciolini.pdf
accesso aperto
Dimensione
86.53 MB
Formato
Adobe PDF
|
86.53 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/97797
URN:NBN:IT:UNIVPM-97797