GPU acceleration and numerical optimization for SPH codes

Lipari, G. (Speaker)

Activity: Talk or presentationTalk or presentation at a workshop, seminar, course or other meeting

Description

GPUs stands for (general-purpose) graphic processing units and SPH for smoothed particle hydrodynamics. SPH is a meshless computational method for Lagrangian (fluid) mechanics suited for simulating bodies with highly-deformable, even fragmented interfaces. Violent-impact flows like slamming gravity waves or colliding galaxies are signature applications. In brief, particles represent volume portions whose individual motion is determined by the interactions within a delimited neighbourhood*. On the downside, advanced simulations require large numbers of particles, say O(10­8). To that end, GPU parallelism effectively accelerates the definition of neighbourhoods and interactions, making compute-intensive simulations feasible. This talk outlines challenges and promises of the namesake 2-year R&D collaboration between the Numerical Analysis group and the Shell Technology Centre in Bangalore, India*. This project has been stimulated by the opportunity to port special functionalities of an existing in-house SPH code with CPU parallelism into a host code with GPU parallelism. The host candidates are open-source solvers developed by either consortia or research communities. On the one hand, the project outcomes serve operational needs, whereby seamless integration in an extant workflow is a key requirement. On the other hand, three knowledge domains need to be invoked and are interwoven horizontally: physics (fluid dynamics), mathematics (formulations and algorithms), computer science (programming, building, execution).
Period5 Nov 2018
Held atDelft Institute of Applied Mathematics
Degree of RecognitionLocal

Keywords

  • SPH
  • Smoothed Particle Hydrodynamics
  • GPU
  • hardware acceleration
  • Scientific Computing