Parallel algorithms for computational fluid dynamics on unstructured meshes

Tesis doctoral de Ricard Borrell Pol

Direct numerical simulation (dns) of complex flows is currently an utopia for most of industrial applications because computational requirements are too high. For a given flow, the gap between the required and the available computing resources is covered by modeling/simplifying of some terms of the original equations. On the other hand, the continuous growth of the computing power of modern supercomputers contributes to reduce this gap, reducing hence the unresolved physics that need to be attempted with approximated models. This growth, widely relies on parallel computing technologies. However, getting the expected performance from new complex computing systems is becoming more and more difficult, and therefore part of the cfd research is focused on this goal. Regarding to it, some contributions are presented in this thesis. The first objective was to contribute to the development of a general purpose multi-physics cfd code. Referred to as termofluids (tf). Tf is programmed following the object oriented paradigm and designed to run in modern parallel computing systems. It is also intensively involved in many different projects ranging from basic research to industry applications. Besides, one of the strengths of tf is its good parallel performance demonstrated in several supercomputers. In the context of this thesis, the work was focused on the development of two of the most basic libraries that compose tf: i) the basic objects library (bol), which is a parallel unstructured cfd application programming interface, on the top of which the rest of libraries that compose tf are written, ii) the linear solvers library (lsl) containing many different algorithms to solve the linear systems arising from the discretization of the equations. The first chapter of this thesis contains the main ideas underlying the design and the implementation of the bol and lsl libraries, together with some examples and some industrial applications. A detailed description of some application-specific linear solvers included in the lsl is carried out in the following chapters. In the second chapter, a parallel direct poisson solver restricted to problems with one uniform periodic direction is presented. The poisson equation is solved, at least, once per time-step when modeling incompressible flows, becoming one of the most time consuming and difficult to parallelize parts of the code. The solver here proposed is a combination of a direct schur-complement based decomposition (dsd) and a fourier diagonalization. The latter decomposes the original system into a set of mutually independent 2d sub-systems which are solved by means of the dsd algorithm. Since no restrictions are imposed in the non-periodic directions, the overall algorithm is well-suited for solving problems discretized on extruded 2d unstructured meshes. The scalability of the solver has been successfully tested using up to 8192 cpu cores for meshes with up to 10¿ grid points. In the last chapter, a solver for the boltzmann transport equation (bte) is presented. It can be used to solve radiation phenomena interacting with flows. The solver is based on the discrete ordinates method and can be applied to unstructured discretizations. The flux for each angular ordinate is swept across the computational grid, within a source iteration loop that accounts for the coupling between the different ordinates. The sequential nature of the sweep process makes the parallelization of the overall algorithm the most challenging aspect. Several parallel sweep algorithms, which represent different options of interleaving communications and calculations, are analyzed. One of the heuristics proposed consistently stands out as the best option in all the situations analyzed. With this algorithm, good scalability results have been achieved regarding both weak and strong speedup tests with up to 2560 cpus.

 

Datos académicos de la tesis doctoral «Parallel algorithms for computational fluid dynamics on unstructured meshes«

  • Título de la tesis:  Parallel algorithms for computational fluid dynamics on unstructured meshes
  • Autor:  Ricard Borrell Pol
  • Universidad:  Politécnica de catalunya
  • Fecha de lectura de la tesis:  31/10/2012

 

Dirección y tribunal

  • Director de la tesis
    • Assensi Oliva Llena
  • Tribunal
    • Presidente del tribunal: Antonio Pascau benito
    • roeland Verstappen (vocal)
    • (vocal)
    • (vocal)

 

Deja un comentario

Tu dirección de correo electrónico no será publicada. Los campos obligatorios están marcados con *

Scroll al inicio