Abstract

The simulation of fluid flow on rectangular grids using a discretized version of the Navier Stokes Equations for incompressible fluid flow can be simultaneously described as an aesthetically pleasing and computationally intensive embarrassingly parallel problem.

Ideally, the aesthetics of the fluid simulation should, given some set of parameters, feel natural despite the synthetic nature of the underlying grids. This natural feel, paramount to the success of the system, should fool a person into believing that they are interacting with a real fluid.

The number of calculations and data accesses increases with the number of cells present in the rectangular grid upon which the fluid is simulated. An increased number of calculations are required for augmented accuracy, different external forces, and additional dimensions. Since it is a trivial task to increase the complexity of the simulation, interactivity becomes a challenge of balancing accuracy, stability, and detail against speed of execution.

A simple solution is to throw more processing power through increased instruction execution speeds or additional cores. Throwing additional cores at the problem strains the memory bus making it the point that slows down the simulation. Therefore for a given algorithm, respecting data locality and processor peculiarities can be used to minimize execution times.

This document introduces a means of caching corrected velocity fields, a task scheduler that attempts to maximize the usage of the cache on multi-core processors, and a na\"ive compression algorithm based on run-length encoding.