Performance Trade-offs of High Order Meshless Approximation on Distributed Memory Systems

Meshless methods approximate operators in a specific node as a weighted sum of values in its neighbours. Higher order approximations of derivatives provide more accurate solutions with better convergence characteristics, but they come at the cost of including more neighbours. On the accuracy-per-compute time basis we know that increasing the approximation order is beneficial for a shared memory computer, but there is additional communication overhead when problems become too large and we have to resort to distributed memory systems. Meshless nodes are divided between systems in spatially coherent subdomains with approximations at their edges requiring neighbouring value exchange. Performance optimization is then a balancing act between minimizing the required number of communicated neighbours by lowering the approximation order or increasing it to enable faster convergence. We use the radial basis function-generated finite difference method (RBF-FD) to approximate the derivatives that we use to solve the Poisson equation with an explicit iterative scheme. Intersystem communication is provided by Open MPI, while OpenMP is used for intra-system parallelisation. We perform the analysis on a homogenous CPU-based cluster where we examine the behaviour and attempt to determine the optimal parameterisation with the goal of minimizing the computational time to reach a desired accuracy.