Parallel architectures for fluid flow simulation

Abstract

The computational requirements for accurate simulations of high Reynolds number flows are massive. Even when extrapolating technologies and parallel architectures, it is unclear whether we could ever assemble enough computational resources to solve the unapproximated Navier-Stokes equations in regimes of realistic Reynolds numbers. One thing is clear, machines dedicated to directly solving turbulent flow problems must efficiently execute complex, irregular algorithms and support rapid experimentation with new techniques in the parallel setting. Fundamentally, these machines must be general purpose and programmable. We present these challenges and outline some possible solutions to fundamental issues in general purpose parallelism. We should also take seriously the possibility that solving Navier-Stokes is not the best approach to full turbulence fluid flow simulation. Instead, direct simulations of microdynamics on rectangular lattices may be far more tractable. Building on recent theoretical breakthroughs, it now appears that Boolean microdynamic systems can be constructed whose mean statistics are precisely those of a real fluid. We show how this approach, as measured in bit transitions per fluid volume update, is fundamentally orders of magnitude more efficient than Navier-Stokes solvers and how we might augment the arithmetic units of current processors to exploit this fact.