An update on tracking point-particles in homogeneous turbulence on heterogeneous (GPU) architectures

Because of the degree of detail readily available, direct numerical simulation (DNS) is an important tool for the study of the motion of fluid particles as well as small particles with inertia in turbulence in canonical geometries amenable to advanced computation.  Criteria for physical fidelity of the simulation results include high Reynolds number, particle population count, and the accuracy of interpolation schemes needed to obtain velocity and other variables following the particle trajectories. We have constructed and applied highly scalable GPU algorithms where, as for the Eulerian velocity field, most operations are performed on GPUs with minimal needs for copying between the host and the device.  For inertial point particles the numerical integration in time incurs significant extra cost only when considering the added-mass (buoyancy) term which requires computing the fluid acceleration at the particle position.  The code has been implemented on both AMD and NVIDIA platforms with different hardware characteristics.  We also discuss briefly what can be learned from animations of particle trajectories originating from localized zones of large velocity gradients.