HPC SPH/N-body

High Performance Computing Smooth(ed) Particle Hydrodynamics/N-body simulations.

Aim

Overall aim is to write a Multi-CPU and/or Multi-GPU (thus HPC) SPH code including self-gravity targeting distributed memory systems.

SPH includes short-range forces and self-gravity or generally gravity corresponds to long-range forces, wherefore solving the N-body problem is a milestone to achieve a SPH code including self-gravity.

Primary aim is a proof of concept.

Challenges

The problem is not trivially parallelizable, especially regarding

• distributed memory
• load balancing

since forces act over the whole domain.

In an abstract way the problem can be summarized to:
reduce communication amount (performance) vs. reduce amount of (locally required) memory

Some consequences:

• Brute-force self-gravity not applicable
• Collecting/Gathering information on one process need to be avoided at all costs

Parallelization

• The Message Passing Interface (MPI) is used for the internode (and currently also for the intranode) communication
• CUDA is used for GPGPU programming targeting NVIDIA GPUs
• OpenMP could be used in the future for intranode optimizations

Implementation

Self-gravity is implemented via the Barnes-Hut method, thus using an Octree, which can be used for the Neighborhood-search as well, enabling SPH.

Multi-CPU

See MultiCPU for more information.

Current status: Multi-CPU N-body & basic SPH working!

Multi-GPU

See MultiGPU for more information.

Current status: Multi-GPU N-body working, but not yet extended to SPH!