INTEC   05402
INSTITUTO DE DESARROLLO TECNOLOGICO PARA LA INDUSTRIA QUIMICA
Unidad Ejecutora - UE
congresos y reuniones científicas
Título:
A FFT Preconditioning Technique for the Solution of Incompressible Flow with Fractional Step Methods on Graphic Processing Units
Autor/es:
MARIO A. STORTI; RODRIGO R. PAZ; LISANDRO D. DALCIN; SANTIAGO D. COSTARELLI
Lugar:
Buenos Aires
Reunión:
Congreso; MECOM del Bicentenario; 2010
Institución organizadora:
Asociación Argentina de Mecánica Computacional
Resumen:
The resolution of Computational Fluid Dynamics (CFD) problems on
Graphic Processing Units (GPUs) requires of specialized algorithms due
to the particular hardware architecture of these devices. Algorithms
that fall in the category of cellular automata (CA) are the best fitted,
for instance explicit Finite Volume or Finite Element methods. But in
the case of incompressible flow it is not possible to develop a pure
explicit algorithm, due to the essentially non-local character of the
incompressibility condition. In this case the algorithms that are closer
to an explicit approach, are segregated algorithms, like the Fractional
Step Method. In this algorithms the more time consuming stage is
(asymptotically for large problems) the solution of the Poissons
equation for pressure. A common choice for its solution is the IOP
(Iterated Orthogonal Projection) method, which requires a series of
solutions on the complete mesh. In this work a variant of the IOP,
called Accelerated Global Preconditioning (AGP), is proposed. It is
based on using a Preconditioned Conjugate Gradient (which is an
accelerated iterative method, in contrast with the stationary scheme
used in IOP) for the pressure on the fluid, and preconditioning with the
solution on the global domain (fluid and solid). Of course, solving the
problem on the global domain represents more computational work than
solving the problem only in the fluid, but this can be faster in a
structured mesh if a fast solvers as Multigrid or Fast Fourier Transform
(FFT) is used. The main advantage of AGP over IOP is that it is an
accelerated solver, whereas the IOP is stationary. In addition AGP
iterates only on pressure, whereas IOP iterates on both pressure and
velocity.