Parallel fluid dynamics for the film and animation industries

Abstract

The creation of automated fluid effects for film and media using computer simulations is popular, as artist time is reduced and greater realism can be achieved through the use of numerical simulation of physical equations. The fluid effects in today’s films and animations have large scenes with high detail requirements. With these requirements, the time taken by such automated approaches is large. To solve this, cluster environments making use of hundreds or more CPUs have been used. This overcomes the processing power and memory limitations of a single computer and allows very large scenes to be created. One of the newer methods for fluid simulation is the Lattice Boltzmann Method (LBM). This is a cellular automata type of algorithm, which parallelizes easily. An important part of the process of parallelization is load balancing; the distribution of computation amongst the available computing resources in the cluster. To date, the parallelization of the Lattice Boltzmann method only makes use of static load balancing. Instead, it is possible to make use of dynamic load balancing, which adjusts the computation distribution as the simulation progresses. Here, we investigate the use of the LBM in conjunction with a Volume of Fluid (VOF) surface representation in a parallel environment with the aim of producing large scale scenes for the film and animation industries. The VOF method tracks mass exchange between cells of the LBM. In particular, we implement the new dynamic load balancing algorithm to improve the efficiency of the fluid simulation using this method. Fluid scenes from films and animations have two important requirements: the amount of detail and the spatial resolution of the fluid. These aspects of the VOF LBM are explored by considering the time for scene creation using a single and multi-CPU implementation of the method. The scalability of the method is studied by plotting the run time, speedup and efficiency of scene creation against the number of CPUs. From such plots, an estimate is obtained of the feasibility of creating scenes of a giving level of detail. Such estimates enable the recommendation of architectures for creation of specific scenes. Using a parallel implementation of the VOF LBM method we successfully create large scenes with great detail. In general, considering the significant amounts of communication required for the parallel method, it is shown to scale well, favouring scenes with greater detail. The scalability studies show that the new dynamic load balancing algorithm improves the efficiency of the parallel implementation, but only when using lower number of CPUs. In fact, for larger number of CPUs, the dynamic algorithm reduces the efficiency. We hypothesise the latter effect can be removed by making using of centralized load balancing decision instead of the current decentralized approach. The use of a cluster comprising of 200 CPUs is recommended for the production of large scenes of a grid size 6003 in a reasonable time frame.