3D Lattice Boltzmann Magneto-hydrodynamics(LBMHD3D)....pptVIP

* 3D Lattice Boltzmann Magneto-hydrodynamics (LBMHD3D) Sam Williams1,2, Jonathan Carter2, Leonid Oliker2, John Shalf2, Katherine Yelick1,2 1University of California Berkeley 2Lawrence Berkeley National Lab samw@cs.berkeley.edu October 26, 2006 Previous Cell Work Lattice Methods LBMHD Implementation Performance Outline Previous Cell Work Sparse Matrix and Structured Grid PDEs Double precision implementations Cell showed significant promise for structured grids, and did very well on sparse matrix codes. Single precision structured grid on cell was ~30x better than nearest competitor SpMV performance is matrix dependent (average shown) 18.0 4.4 5.6 14.63 Peak GFLOP/s 1.14 3.91 X1E 0.36 0.57 Opteron 0.36 1.19 Itanium2 2.66 7.16 Cell SpMV GFLOP/s Structured Grid GFLOP/s Architecture Quick Introduction to Lattice Methods and LBMHD Lattice Methods Lattice Boltzmann models are an alternative to top-down, e.g. Navier-Stokes and bottom-up, e.g. molecular dynamics algorithms, approaches Embedded higher dimensional kinetic phase space Divide space into a lattice At each grid point, particles in discrete number of velocity states Recovery macroscopic quantities from discrete components Lattice Methods (example) 2D lattice maintains up to 9 doubles (including a rest particle) per grid point instead of just a single scalar. To update one grid point (all lattice components), one needs a single lattice component from each of its neighbors Update all grid points within the lattice each time step 3D Lattice Rest point (lattice component 26) 12 edges (components 0-11) 8 corners (components 12-19) 6 faces (components 20-25) Total of 27 components, and 26 neighbors 2 11 7 1 8 4 9 6 0 10 5 3 14 18 19 17 13 12 15 16 22 25 21 20 23 24 LBMHD3D Navier-Stokes equations + Maxwell’s equations. Simulates high temperature plasmas in astrophysics and magnetic fusion Implemented in Double Precision Low to moderate Reynolds number LBMHD3D Originally developed by George Vahala @ Col