[ HPF Home | Versions | Compilers | Projects | Publications | Applications | Benchmarks | Events | Contact ]

---

SPH

Back to Applications

Erik Schnetter
Theoretische Astrophysik
Universitaet Tuebingen
Germany
Web: http://www.tat.physik.uni-tuebingen.de
Email: schnetter@tat.physik.uni-tuebingen.de

The application uses Smoothed Particle Hydrodynamics (SPH) to simulate compressible flows. SPH is a Lagrangian particle method that is widely used in astrophysics. It is algorithmically (but not physically!) very similar to Molecular Dynamics (aka MD).

Number of Lines of Code: 1000

Target Platforms and HPF Compilers Used:

Linux PC cluster (PGHPF) Cray T3E (PGHPF)

Coding Styles (data decompositions, computational methods):

The list of particles and a list of interacting particle pairs are both stored as distributed arrays. There is no explicit domain decomposition.

Extrinsic Interfaces Used (and reasons):

We use extrinsic (hpf_local) to flag routines that don't require communication.

Performance Information, if Available (including any possible comparisons to MPI and/or OpenMP):

We also have a MPI version of the code
that uses explicit domain decomposition
and has been hand-tuned for memory
consumption. The (yet untuned) HPF
version uses a lot more memory and is
also several times slower at finding
interacting particles. Evaluating
interactions happens at the same speed,
therefore the overall computing times
are comparable. The HPF version scales
better with large numbers of processors.

Please comment on any aspects of the application that might be interesting, including any problems using HPF effectively:

SPH is usually limited by bookkeeping and by memory bandwith; floating point performance is less important. We did not define a data type "particle" because we found that the HPF compilers have trouble with distributed arrays of derived types. The current code evaluates only the RHS of the equations without integration in time (which should be trivial to add).