| Name: fluidanimate |
| |
| Description: |
| |
| Smoothed Particle Hydrodynamics (SPH) is a computational method for simulating |
| fluid flows. This benchmark application computes one time-step of a liquid |
| simulation which solves the Navier-Stokes equations using SPH. The liquid is |
| represented by a collection of Newtonian particles which interact with each |
| other. These interactions are short-range which enables the use of a uniform |
| grid to accelerate the determination of which pairs of particles interact. After |
| computing the density of the fluid at each particles position, the acceleration |
| of each particle due to gravity, fluid pressure and fluid viscosity is computed. |
| Next, collision detection is performed on six, axis-aligned planes (defining a |
| box) and collision response is done with the penalty method. Finally, numerical |
| time integration, using the Leapfrog method, calculates the new position and |
| velocity of each particle based on its acceleration. |
| |
| About 90% of execution time is spent computing particle densities and forces, |
| about 7% is spent sorting particles into the uniform grid, and the remaining |
| time is spent handling collisions and performing time integration. |
| |
| Parallelization: |
| |
| The parallelization algorithm provided is based on spatial partitioning. As |
| mentioned above, particles are sorted spatially into a uniform grid which |
| covers the entire simulation domain. This grid is evenly partitioned along cell |
| boundaries using axis-aligned splitting planes to produce subgrids, one per |
| thread. Since particles which reside in adjacent cells may interact with each |
| other, multiple threads may need to update the cells which lie on sub-grid |
| boundaries. |
| |
| Each grid cell has a flag which indicates whether or not it lies on a subgrid |
| boundary. Whenever a particle in a boundary cell needs to be updated, |
| fine-grain locks are used for that update. Locks are not used when updating |
| particles away from boundaries since these updates will never occur |
| concurrently. |
| |
| ======================================= |
| Programming Languages & Libraries |
| |
| C++ and pthread is used in this benchmark. |
| |
| ======================================= |
| System Requirements: |
| |
| 1) Intel(R) C++ Compiler: version 9.0 or higher |
| 2) GNU gcc/g++: version 3.3 or higher |
| 3) sed: version 4.0.9 or higher recommended. |
| At least 200 MBytes memory is needed for the benchmark. |
| ======================================= |
| Input/Output: |
| |
| The dataset of the benchmark contains the data of a collection of Newtonian |
| particles which interact with each other. The program calculates the new |
| position and velocity of each particle as the output. |
| |
| ======================================= |
| Characteristics: |
| |
| (1) Hotspot |
| 90% of execution time is spent computing particle densities and forces. |
| "ComputeDensitiesMT", "ComputeDensities2MT" and "ComputeForcesMT" are |
| the main functions. |
| |
| ======================================= |
| Revision History |
| 12/5/06: Richard O. Lee - First version |
| |
| ======================================= |
| Author: Richard O. Lee |
| |
| Acknowledgements: Mikhail Smelyanskiy supervised this work. |
| Adam Kerin packaged code |
| |
| References: This work is based on the following paper: |
| |
| M. MΓΌller, D. Charypar, M. Gross, Particle-Based Fluid Simulation for |
| Interactive Applications, in Proceedings of ACM SIGGRAPH Symposium on Computer |
| Animatio (SCA) 2003, pp 154-159 |
| (http://graphics.ethz.ch/%7Emattmuel/publications/sca03.pdf) |
| |
