Contribution of Python to LMGC90 platform

Transcription

1 Contribution of Python to LMGC90 platform Bagneris M., Dubois F., Iceta D., Martin A. and Mozul R. Laboratoire de Mécanique et Génie Civil Université Montpellier 2 - Centre National de la Recherche Scientifique EuroScipy 2010

2 Summary 1. Software features and applications

3 Purpose Modeling a large collection of 2D/3D objects with interactions and complex mechanical behavior.

4 Purpose Modeling a large collection of 2D/3D objects with interactions and complex mechanical behavior. Some fields of application : Granular material from rheology to structure study the behavior of a collection of bodies, influence of shapes, interaction laws and bulk properties. (Renouf M. et al. 2004)

5 Purpose Modeling a large collection of 2D/3D objects with interactions and complex mechanical behavior. Some fields of application : Granular material from rheology to structure Masonry structures access the stability and the safety of masonry structures, static or dynamic loads, influence of the design pattern and the joint behavior. (Rafiee A. et al. 2008)

6 Purpose Modeling a large collection of 2D/3D objects with interactions and complex mechanical behavior. Some fields of application : Granular material from rheology to structure Masonry structures Fracture of heterogeneous media fracture modeled at the microscopic or the macroscopic scale, from initiation to post-failure, Frictional Cohesive Zone Model. (Perales F. et al. 2005)

7 Purpose Modeling a large collection of 2D/3D objects with interactions and complex mechanical behavior. Some fields of application : Granular material from rheology to structure Masonry structures Fracture of heterogeneous media Rock mass stability rock mass considered as fractured media, study of avalanche on a natural or mining slopes and stability of tunnels. (Rafiee A. et al. 2008)

8 Purpose Modeling a large collection of 2D/3D objects with interactions and complex mechanical behavior. Some fields of application : Granular material from rheology to structure Masonry structures Fracture of heterogeneous media Rock mass stability Multi-physics couplings various physics may be considered (different scales), e.g. thermal coupling, fluid particle interaction, electrical conductivity. (Topin V. et al. 2009)

9 Main features Modeling Object shapes : simple convex primitives (e.g. disks, polygons, spheres, polyhedra), cluster of primitives, general triangulated surfaces, contact detetction may be performed for any combination of shapes.

10 Main features Modeling Object shapes : simple convex primitives (e.g. disks, polygons, spheres, polyhedra), cluster of primitives, general triangulated surfaces, contact detetction may be performed for any combination of shapes. Bulk behaviors : rigid, deformable, by means of the FEM (e.g. elastic, hyperelastic, vicous, plastic), influence of other physics (e.g. thermic, fluid dynamics).

11 Main features Modeling Object shapes : simple convex primitives (e.g. disks, polygons, spheres, polyhedra), cluster of primitives, general triangulated surfaces, contact detetction may be performed for any combination of shapes. Bulk behaviors : rigid, deformable, by means of the FEM (e.g. elastic, hyperelastic, vicous, plastic), influence of other physics (e.g. thermic, fluid dynamics). Interactions laws : a large set available (e.g. frictional contact, cohesion, wire, rod).

12 Main features Modeling Object shapes : simple convex primitives (e.g. disks, polygons, spheres, polyhedra), cluster of primitives, general triangulated surfaces, contact detetction may be performed for any combination of shapes. Bulk behaviors : rigid, deformable, by means of the FEM (e.g. elastic, hyperelastic, vicous, plastic), influence of other physics (e.g. thermic, fluid dynamics). Interactions laws : a large set available (e.g. frictional contact, cohesion, wire, rod). Multi-physics : multiple discretizations available, physic couplings at different scales, interpolation/projection methods.

13 Main features Modeling Object shapes : simple convex primitives (e.g. disks, polygons, spheres, polyhedra), cluster of primitives, general triangulated surfaces, contact detetction may be performed for any combination of shapes. Bulk behaviors : rigid, deformable, by means of the FEM (e.g. elastic, hyperelastic, vicous, plastic), influence of other physics (e.g. thermic, fluid dynamics). Interactions laws : a large set available (e.g. frictional contact, cohesion, wire, rod). Multi-physics : multiple discretizations available, physic couplings at different scales, interpolation/projection methods. Analysis Common method, the Non Smooth Contact Dynamics : non smooth dynamic framework, implicit time integration, implicit contact solvers (e.g. NLGS, GPCP).

14 Main features Modeling Object shapes : simple convex primitives (e.g. disks, polygons, spheres, polyhedra), cluster of primitives, general triangulated surfaces, contact detetction may be performed for any combination of shapes. Bulk behaviors : rigid, deformable, by means of the FEM (e.g. elastic, hyperelastic, vicous, plastic), influence of other physics (e.g. thermic, fluid dynamics). Interactions laws : a large set available (e.g. frictional contact, cohesion, wire, rod). Multi-physics : multiple discretizations available, physic couplings at different scales, interpolation/projection methods. Analysis Common method, the Non Smooth Contact Dynamics : non smooth dynamic framework, implicit time integration, implicit contact solvers (e.g. NLGS, GPCP). Other strategies may be implemented (modular architecture).

15 Preprocessing tool First step before any computation : preprocessing geometry description, bulk/interaction behaviors.

16 Preprocessing tool First step before any computation : preprocessing geometry description, bulk/interaction behaviors. Heavy FORTRAN readable file format dedicated preprocessors needed.

17 Preprocessing tool First step before any computation : preprocessing geometry description, bulk/interaction behaviors. Heavy FORTRAN readable file format dedicated preprocessors needed. Old-fashioned : many monolithic compiled programs no possible interactions no common parts (data structures/files writing)

18 Preprocessing tool First step before any computation : preprocessing geometry description, bulk/interaction behaviors. Heavy FORTRAN readable file format dedicated preprocessors needed. Old-fashioned : many monolithic compiled programs no possible interactions no common parts (data structures/files writing) New-fashioned (Python powered) : a unified generic architecture (object oriented) suitable for objects description embeds all previous prepocessor features (shared libraries) allows complex samples generation

19 Core and Supervisor Second step after preprocessing : computation

20 Core and Supervisor Second step after preprocessing : computation Core written in FORTRAN 90 : well known performances in numerical computing applications, frequently used language in academic software.

21 Core and Supervisor Second step after preprocessing : computation Core written in FORTRAN 90 : well known performances in numerical computing applications, frequently used language in academic software. User interface needs : modularity, flexibility, fine description of all simulation stages scritping language.

22 Core and Supervisor Second step after preprocessing : computation Core written in FORTRAN 90 : well known performances in numerical computing applications, frequently used language in academic software. User interface needs : modularity, flexibility, fine description of all simulation stages scritping language. Python arises as a good solution, since it offers : many control structures, object-oriented programming, various modules (e.g. NumPy, SciPy), and other language interfacing tools.

23 From FORTRAN 90 to Python Old-fashioned : f2py An automatic tool giving access to ALL entities of a module (variables and functions) defining a limited API needs a specific handmade treatment : Core mod_rbdy2.f90 (FORTRAN 90) Wrap wrap_rbdy2.f90 (FORTRAN 90) major issues : management of compiler options, compilers detection. object files compile f2py chipy.so

24 From FORTRAN 90 to Python New-fashioned : SWIG An automatic interface generator for C/C++ libraries, FORTRAN 2003 ISO C Binding feature allows FORTRAN routines to be called as C functions (available in latest releases of compilers, e.g. ifort, gfortran, g95). Core Wrap mod_rbdy2.f90 (FORTRAN 90) wrap_rbdy2.f90 (FORTRAN 2003) wrap_rbdy2.h (handmade) include chipy.i (handmade) numpy.i, docstring compile compile SWIG object files compile chipy.cpp chipy.py link _chipy.so import

25 Parallelization LMGC90 interactions solver is already parallelized using multi-threading thanks to OpenMP.

26 Parallelization LMGC90 interactions solver is already parallelized using multi-threading thanks to OpenMP. To use a cluster architecture, a domain decomposition method is developped, but MPI is needed.

27 Parallelization LMGC90 interactions solver is already parallelized using multi-threading thanks to OpenMP. To use a cluster architecture, a domain decomposition method is developped, but MPI is needed. Proper way : reengineer the core of LMGC90 (i.e in the deepest FORTRAN routines) very time consuming.

28 Parallelization LMGC90 interactions solver is already parallelized using multi-threading thanks to OpenMP. To use a cluster architecture, a domain decomposition method is developped, but MPI is needed. Proper way : reengineer the core of LMGC90 (i.e in the deepest FORTRAN routines) very time consuming. Next idea (Iceta s PhD) : use a Python MPI module (e.g. mpi4py) to implement the domain decomposition method at the highest level (i.e. the Python supervisor) : not efficient because of vector copy between Python and FORTRAN, + : very easy to implement allow to test the method with really low developing cost.

29 Conclusion Improvements thanks to Python : a full preprocessing tool, a flexible user interface, allowing quick implementation of sophisticated methods (e.g. domain decomposition, couplings). Remaining expectation : design a postprocessing tool plotting using MatPlotLib, visualization using ParaView and Mayavi. Open source software : feel free to use or to contribute frederic.dubois@univ-montp2.fr.