There are two computationally intensive tasks commonly used in artistic content creation: simulations and rendering. Oftentimes, these can form severe bottlenecks that may jeopardize the deadlines of film productions; as such, they are both excellent candidates for High-Performance Computing (HPC) applications for the MSC-BW project. This article presents Palander, a prototype for one of these applications, which has been built by our project partners at the Animationsinstitut of the Filmakademie Baden-Wuerttemberg; it is currently being tested within a student project as well as in a collaboration with a local visual effects company.
The following article is also available as pdf.
Palander is a prototype for using a digital content creation (DCC) tool to direct a fluid simulation on an external solver, which can be run on the HLRS Cray supercomputer, Hazel Hen. The name “Palander” is a combination of the tools used to achieve this: Palabos is the external fluid solver framework, and Blender is the DCC tool. Both softwares are open-source, which not only allows them to be freely adapted to this specific use case, but also removes the concern of having to obtain multiple costly licenses for running the software on a distributed system like the Hazel Hen. In addition, all independent files that constitute the Palander prototype are released as open-source; all the necessary files and/or instructions for where to find them are freely available at the following address: https://github.com/FilmakademieRnd/Palander
As a Blender add-on, Palander presents the user with an additional menu full of options used to run the simulation, in addition to a number of parameters extracted automatically from the scene setup. Palander can run a Palabos fluid simulation on the local desktop for quick testing as well as on the Hazel Hen. When either of these functionalities is activated, the prototype automatically transfers all of the information required to run the simulation — from numerical parameters to mesh geometries — and initiates the simulation on the desired platform. The simulation results are automatically meshed and transferred back to Blender for viewing and rendering. Naturally, nothing stops the user from running the same simulation on Blender’s native simulation engine (Elbeem) as well, for further comparison of results.
Benchmark testing has shown that depending on the scene setup, Palander can benefit from a high parallel efficiency as simulation times are kept down by adding a proportional amount of cores to simulations with increasing detail (see Fig. 1). Comparison runs on a desktop computer with a constant amount of processing cores show that an increase in detail causes the time requirement to grow exponentially, until the memory requirement can no longer be fulfilled and the simulation crashes. Memory is not a concern when running on Hazel Hen, since the addition of cores for processing brings in a corresponding amount of new available memory as well.
In rendered images of sample results (see Fig. 2), we see the effect of the increased detail: smaller and smaller features, such as individual droplets, bubbles, and even foam, become distinguishable. Thus, by adding computational power to the simulations, it is possible to bypass the need for e.g. complementary particle simulations often used in conventional production methods to achieve these supplementary effects.
Our ongoing research efforts center around further testing and improving of the prototype to show that it can be used in production-scale projects as well. We are currently working in close collaboration with the Animationsinstitut student group responsible for the FMX 2018 official trailer “Biber im Bart”, with the goal of using Palander fluid simulations in one of their episodes. Likewise, we are cooperating with local visual effects companies in Stuttgart in an effort to compare Palander results to the work made in a professional production environment.