Abstract

Different graph partitioning methods, i.e., linear partioning, parallel hypergraph (PHG) partioning, and two approaches using ParMETIS, are considered to generate an unstructured decomposition of the second-level coarse operator of three-level FROSch (Fast and Robust Overlapping Schwarz) preconditioners in the Trilinos software library. In our context, the parallel hypergraph method shows the most consistent results.

Abstract

Coupled fluid-particle simulations are routinely used in a variety of applications, ranging from respiratory droplet spreading to internal combustion engines, from ink-jet printing to in-flight ice accretion. The efficiency of parallel algorithms to simulate fluid-particle systems is strongly influenced by the different evolution of the flow and the particles dynamics. Indeed, a domain partitioning based on particle workload is possibly sub-optimal in terms of the number of fluid volume elements associated to each process. In this work, an efficient mesh partitioning based on graph representation is implemented. It can handle unstructured hybrid meshes composed by triangles and quadrilaterals in two spatial dimensions, and by tetrahedra, hexahedra, prisms, and pyramids in three dimensions. In order to obtain a domain decomposition to efficiently follow the particle trajectories, a preliminary solution is computed to suitably tag the fluid domain cells. The obtained weights represent the element probabilities to be crossed by particles. The algorithm is implemented using MPI distribute memory environment. The proposed approach is tested against reference cases for the coupled flow-particle simulation of ice accretion over 2D and 3D geometries. Two different cloud droplet impact test cases have been simulated: a NACA 0012 wing section and a NACA 64A008 swept horizontal tail. The computed collection efficiency compares fairly well with reference numerical and experimental data. The parallel efficiency of the algorithm is verified on a distributed memory cluster.

Abstract

Abstract

Particle tracking within an underlying flow field is routinely used to analyse both industrial processes and natural phenomena. In a computer code running on a distributed-memory architecture, the different behaviour of fluid-particle systems must be taken into account to properly balance element-particle subdivision among processes. In unsteady simulations, the parallel efficiency is even more critical because it changes over time. Another challenging aspect of a scalable implementation is the initial particle location due to the arbitrary shapes of each subdomain. In this work, an innovative parallel ray tracing particle location algorithm and a two-constrained domain subdivision are presented. The former takes advantage of a global identifier for each particle, resulting in a significant reduction of the overall communication among processes. The latter is designed to mitigate the load unbalance in the particles evolution while maintaining an equal element distribution. A preliminary particle simulation is performed to tag the cells and compute a weight proportional to the probability to be crossed. The algorithm is implemented using MPI distribute memory environment. A cloud droplet impact test case starting from an unsteady flow around a 3D cylinder has been simulated to evaluate the code performances. The tagging technique results in a computational time reduction of up to 78% and a speed up factor improvement of 44% with respect to the common flow-based domain subdivision. The overall scalability is equal to 1.55 doubling the number of cores.