VPAC Facility Introduction by gvwf5Cc4

VIEWS: 0 PAGES: 20

									Bill Appelbe, bill@vpac.org
                              1
THE SNARK PROJECT

Outline
Why do we need another solver?
Design of Snark
Implementation timetable




                                  2
Why Another Solver

 No solver meets the requirements we
  have
 The fundamental requirement is that
  the solver be
  1. Adapt to a wide range problems
  2. Fast (parallel)
  3. Accurate


                                        3
Why Another Solver

 Range of problems:
   Geological CFD
    1. Mantle convection
    2. Regional crustal deformation
        Ore body formation
    3. Tectonic modeling
    4. Multiphase flows
   Engineering problems with complex
    rheologies

                                        4
Design
Choices

First rule
  Decouple, decouple, decouple
     o Setup of initial conditions
     o Setup of grid (meshing)
        • And regriding
     o Domain decomposition (for scaleable
       parallelism)
     o Solver
        • Should be capable of handing arbitrary material
          properties
     o Viewing of output
                                                            5
Design
Choices

                                         Rheology
                                         specification
User input                               Compiler
or another grid format

        (Re)-Grid            Domain
        Generator        (re)Decompose
                                           Solver
                                           Engine

                           Viewer
                                                         6
Features of Snark

 Snark supports the functionality of the
   ELLIPSIS solver (Moresi) in addition to:
1. new general purpose parallel execution
2. an algebraic multigrid (AMG) technique
   coupled with the particle-in-cell (PIC)
   method
3. Decoupling of equations from generic
   MG/PIC
   Exploit C++ template functions
4. Generic boundary conditions and grid
   generator
                                              7
Basic Components of Snark

 The Snark system consists of five portable
  subsystems:
  The mesh generator(MGen);
  An interface mesh viewer;
  The Snark FEM engine;
  Resulting data analyzer; and
  “Compiler” for rheology specifications
 By using well defined API’s, these
  components can be “plug and play”

                                               8
The Snark Mesh Generator

 Generating the FE mesh is the most
  difficult part of the modeling process;
 Mesh design is the most critical decision in
  FEM+AMG modeling;
 The mesh generation needs to reflect the
  later AMG as well as the parallel
  algorithms;
 The design should consider plugins for
  versions in the future.


                                                 9
The Mesh_Generator(MGen):

 2D and 3D spatial dimensions for the
  problem domain;
 Triangular mesh for 2D and tetrahedral
  mesh for 3D problems (plug in);
 Linear and quadratic (plug in) elements;
 Structured mesh;
 Isotropic (uniform) mesh;
 Adaptive mesh refinement(plug in);

                                             10
How to Develop the MGen?

Based on our comprehensive review of
 existing mesh generators;
Adapt the features from the existing
 mesh generators, which are suitable
 to MGen;
The VPAC in-house expertise in
 science and engineering as well as
 software engineering;

                                    11
The First Version: Snark_v1

 The focus of Snark_v1 is to apply the PIC
  and AMG into the FEM formulation;
 Steady state field problems, describing
  using linear elliptic PDEs; such as
  Elasticity;
  Heat conduction;
  Seepage through porous media;
  Irrotational flow of idea fluids;
  Distribution of electrical (or magnetic)
   potential;etc

                                              12
What’s MG and AMG

 Multigrid (MG) methods are fast linear
  iterative solvers based on multilevel, or
  multiscale paradigm
 Two choices:
  Geometric vs. Algebraic multigrid
 Algebraic multigrid (AMG) methods are any
  multilevel methods where the geometry is not
  used (maybe not available) .


                                              13
Why MG (AMG)

 MG is amongst the fastest solution
  techniques known today;
 MG can treat arbitrary regions and
  boundary conditions;
 MG directly applies to more complicated
  situations;
 MG exhibits a convergence rate that is
  independent of the number of the unknowns
  in the discretized system
 MG is amenable to parallel processing

                                          14
Implementation Timetable

 First step has been a through organized
  survey of the literature on available
  solvers, MG, code, etc.
  As an internal website
 Delivery of beta of Mgen by the end of
  2001
  Couple to the current Ellipsis backend
  For feedback and testing
 Replace current ellipsis engine by an AMG
  engine
  Mid 2002
                                              15
Related Projects

 Platyplus-V2
  A scaleable plate reconstruction tool
  In beta release form after 6mo. development
 A new generic SPH solver
  MNRF and Monash funded
 The solvers operate as a “suite”
 Snark vs Fastflo
  Our aim is performance, parallelism, large
   problems
  Fastflo is aimed at prototyping

                                                 16
VPAC HPC Facility

Compaq AlphaServer SC
  o 32 ES40’s (4 x 833MHz alpha’s, 2G
    Memory)
  o Quadrics interconnect (6μsec latency)
  o 128 processors,
  o 100+ Gigaflops
  o 1.4T shared file
  o server, dual
  o DLT tapes

                                            17
VPAC HPC Facilities

  VPAC “owns” 5% of the cycles of RMIT’s
   Virtual Reality Centre
   o For demonstration projects, PR
   o Tools, training, and support for
     visualisation




                                            18
Contact Us


 Victorian Partnership for Advanced Computing
 Website: http://www.vpac.org




                                           19
Acknowledgement


 Bill Appelbe, Feng Wang, Dev Behera
Victorian Partnership for Advanced Computing
   PO Box 211, Carlton South, Victoria 3053

   Louis Moresi and Hans Mühlhaus
       CSIRO Exploration and Mining,
      PO Box 437, Nedlands, WA 6009

                                           20

								
To top