SLIDE 1
PySPH: A Python framework for SPH
Prabhu Ramachandran Chandrashekhar P . Kaushik
Department of Aerospace Engineering and Department of Computer Science and Engineering IIT Bombay
PySPH: A Python framework for SPH Prabhu Ramachandran - - PowerPoint PPT Presentation
PySPH: A Python framework for SPH Prabhu Ramachandran - - PowerPoint PPT Presentation
PySPH: A Python framework for SPH Prabhu Ramachandran Chandrashekhar P . Kaushik Department of Aerospace Engineering and Department of Computer Science and Engineering IIT Bombay SciPy 2010 Austin, TX, June 30 July 1, 2010 Outline
SLIDE 2
SLIDE 3
Outline Introduction PySPH Architecture
SLIDE 4
SPH examples
SLIDE 5
SPH examples
SLIDE 6
3D
SLIDE 7
Parallel?
SLIDE 8
Smoothed Particle Hydrodynamics
◮ Particle based, Lagrangian ◮ Gingold and Monaghan (1977),
Lucy (1977)
◮ PDEs ◮ Complex problems ◮ Moving geometries ◮ Free surface problems ◮ . . .
SLIDE 9
Smoothed Particle Hydrodynamics
◮ Particle based, Lagrangian ◮ Gingold and Monaghan (1977),
Lucy (1977)
◮ PDEs ◮ Complex problems ◮ Moving geometries ◮ Free surface problems ◮ . . .
SLIDE 10
SPH basics
◮ The SPH approximation
f(r) =
- f(r ′)W(r − r ′, h)dr ′
◮ W(r − r ′, h): kernel, compact support ◮ h: size of the kernel
SLIDE 11
SPH basics
SLIDE 12
SPH basics
neighbors kh Nearest
SLIDE 13
SPH basics . . .
◮ Derivatives: transferred to the kernel ◮ Lagrangian ◮ PDE → ODE
SLIDE 14
Scale-up for larger problems requires parallelization
SLIDE 15
Motivation Reproducible and open research
SLIDE 16
Outline Introduction PySPH Architecture
SLIDE 17
PySPH
◮ SPH + Parallel + Python ◮ pysph.googlecode.com ◮ Open Source (BSD)
SLIDE 18
Requirements and installation
◮ Python, setuptools ◮ numpy, Cython-0.12,
mpi4py-1.2
◮ Mayavi-3.x (optional) ◮ Standard Python package
install
SLIDE 19
High-level solver outline
SLIDE 20
Serial Dam break
solver = FSFSolver ( time_step =0.0001 , total_simulation_time =10. , kernel=CubicSpline2D ( ) ) # create the two e n t i t i e s . dam_wall = Solid (name= ’ dam_wall ’ ) dam_fluid = Fluid (name= ’ dam_fluid ’ ) # The p a r t i c l e s f o r the wall . rg = RectangleGenerator ( . . . ) dam_wall . add_particles ( rg . g e t _ p a r t i c l e s ( ) ) solver . add_entity ( dam_wall )
SLIDE 21
Serial Dam break . . .
# P a r t i c l e s f o r the l e f t column of f l u i d . rg = RectangleGenerator ( . . . ) dam_fluid . add_particles ( rg . g e t _ p a r t i c l e s ( solver . add_entity ( dam_fluid ) # s t a r t the solver . solver . solve ( )
SLIDE 22
Parallel Dam break
solver = ParallelFSFSolver ( time_step =0.0001 , total_simulation_time =10. , kernel=CubicSpline2D ( ) ) # Load p a r t i c l e s in proc with rank 0.
SLIDE 23
Outline Introduction PySPH Architecture
SLIDE 24
Software architecture
◮ Particle kernel ◮ SPH kernel ◮ Solver framework ◮ Serial and Parallel solvers
SLIDE 25
Software architecture
Solver Framework
NNPS SPH Cell Manager Particle Arrays
Solvers
(Python & Cython) (Cython) (Cython) (Cython) (Cython)
C−Arrays
(Cython)
SLIDE 26
Particle kernel
◮ C-arrays: numpy-like but faster and
resizable
◮ ParticleArray: arrays of properties ◮ NNPS (Nearest Neighbor Particle
Search)
◮ Cell, CellManager ◮ Caching
SLIDE 27
SPH kernel
◮ SPH particle interaction: interaction
between 2 SPH particles
◮ SPH summation: interaction of all
particles
SLIDE 28
Solver framework
◮ Entities (made of particles) ◮ Solver components (do various things) ◮ Component manager (checks
property requirements)
◮ Integrator ◮ Basic Solver
SLIDE 29
Parallel solver
◮ Parallel cell manager ◮ Parallel solver components
SLIDE 30
Parallel solver
CellManager Processor 1 Processor 2 Particles Solver components Particles Solver components Parallel case ParallelCellManager ParallelCellManager Serial case Particles Solver components
SLIDE 31
Challenges
◮ Particles move all over ◮ Fixed partitioning will not work ◮ Scale up ◮ Automatic load balancing?
SLIDE 32
Terminology: Cell Cells Particles
SLIDE 33
Terminology: Region
Region 2 P1 P2 Region 1
SLIDE 34
P2
O (−1,−1,0) (1,1,0) (1,−1,0) (2,0,0)
P1 P3 P4
Terminology: Domain decomposition
SLIDE 35
Approach
◮ Domain decomposition: Cells ◮ Cells: dynamically created/destroyed ◮ Processors manage Regions ◮ Cells moved to balance load
SLIDE 36
Load Balancing
Donate Cells
◮ Boundary cells ◮ Cells with least number of local
neighbors
SLIDE 37
Load balancing
SLIDE 38
Efficiency
SLIDE 39
Efficiency
SLIDE 40
Current capabilities
◮ Fully automatic, load balanced,
parallel framework
◮ Relatively easy to script ◮ Good performance ◮ Relatively easy to extend ◮ Free surface flows
SLIDE 41
Immediate improvements
◮ Solver framework redesign ◮ More documentation ◮ Reduce dependency on TVTK for
easier installation
◮ Improved testing on various platforms ◮ Gas dynamics (coming soon) ◮ Solid mechanics (next year)
SLIDE 42