Investigation of possible acceleration of OpenFOAM With the use of FPGA

Abstract

Increasing software performance makes it possible to compute problems consisting larger data sets, which corresponds to more accurate real life simulations. However, there is always strive to achieve even faster computations is always desirable due to profiting from reduced time. By implementing hardware designed to perform a specific task can accelerate computations as well as it is power efficient. OpenFOAM is a computational fluid dynamics tool able to simulate real life problems such as thermodynamics. The authors got the task to investigate if OpenFOAM is suited to be accelerated with an FPGA. During the investigation a model was built from an existing model and after profiling the computations, the Gauss-Seidel smoother where chosen to be implemented on a FPGA. Several attempts were made to understand the arithmetic function of the smoother. The implementation made for it is actually slower than the regular CPU based software. One of the reasons is that the PCIe cannot transfer data fast enough to the FPGA, to resolve this, another smoother have to be implemented that get the same result as Gauss-Seidel. Such as Jacobi's method which can be even more parallelized. Other possibilities include finding different matrix compression method which is more beneficial for the task or even use GPU:s as accelerators