The aim of this project is to optimize a numerical simulation of Computational Fluid Dynamics (CFD) and to make it as scalable as possible. The base code is voluntarily "unoptomized" and contains errors that will need to be fixed through the use of debugging tools like GDB or Valgrind. Afterwards, the goal will be to identify and remove the different contentions in the program in order to achieve peak performance. To that extent, we shall use tracing and profiling tools such as MAQAO, Tau or Interpol. These will help us modify the base implementation and obtain a hybrid code that uses both MPI and multi-threading (e.g. through automatic parallelization with OpenMP).
The provided code aims at simulating a Karman vortex street. We consider a fluid flowing down a 2D tube in which we place a round-shaped obstacle. This simulates, in principle, the setup we would have in a wind tunnel. As for the resolution, we use the Lattice Boltzmann Method (LBM), a relatively recent technique which consists in replacing the Navier-Stokes equations with a discretization of the Boltzmann equation in order to simulate the complex behavior of fluids using streaming and collision (relaxation) processes. This method is particularly well- suited for use in High Performance Computing (HPC) as it is highly parallelizable, which efficiently leverages the architecture of modern multi-core CPUs (Central Processing Units) and GPGPUs (General Purpose Graphical Processing Units).
The following dependencies are needed to run the programs:
- an MPI implementation and its patched C compiler (tested with
OpenMPIv4.1.2,MPICHv4.0.2 andMPCv4.1.0); - GNU Make
Start by cloning this repository locally:
git clone https://github.com/dssgabriel/lattice-boltzmann-method.git
The directory is composed of several sub-directories, each containing an improved version of the simulation. Inside these, simply use the provided Makefiles to compile and/or run the programs:
make build
make run
Other subcommands are available in the Makefiles, such as make gif which generates a GIF of the simulation.
Note: the
make tracecommand requires you to have the Interpol profiler installed on your machine.