Hierarchical Hybrid Grids

Transcription

1 Hierarchical Hybrid Grids IDK Summer School 2012 Björn Gmeiner, Ulrich Rüde July, 2012

2 Contents Mantle convection Hierarchical Hybrid Grids Smoothers Geometric approximation Performance modeling 2

3 Mantle convection 3

4 Simulation of seismic velocities Fig. by Schuberth and Moder - Simulation with TERRA (LMU) 4

5 Boussinesq model for mantle convection problems derived from the equations for balance of forces, conservation of mass and energy: T t (2ηɛ(u)) + p = ρ(t )g, u = 0, + u T (κ T ) = γ. u velocity p dynamic pressure T temperature ν viscosity of the material ɛ(u) = 1 2 ( u + ( u)t ) strain rate tensor ρ density κ, γ, g thermal conductivity, heat sources, gravity vector 5

6 Discretization Temporal for the temperature (explicit): Modified Euler BDF-2 scheme Spatial with FE for the Stokes system: Add an artificial compressibility term Choose a pair of FE spaces that satisfy the LBB condition: Q d q+1 xq q, (q 1) - Taylor-Hood elements Q d q+1 xp q, (q 0) - discontinuous elements 6

7 Weak form: Stokes equation (ɛ(ϕ u i ), 2ηɛ(u n h )) ( ϕu i, p n h ) = (ϕu i, ρ(t n h )g), (ϕ p l, un h ) = 0. In matrix notation (saddle point structure with zero diagonal block): ( ) ( ) ( ) A B T U n F n B 0 P n = 0 Pressure correction approach: ( ) ( ) ( ) A B T U n F n 0 BA 1 B T P n = 0 A ij = (ɛ(ϕ u i ), 2ηɛ(ϕu j )) B ij = (ϕ q i, ϕu j ) F n i = (ϕ u i, ρ(t n )g) 7

8 Boussinesq model for mantle convection problems T t Additional challanges: (2ηɛ(u)) + p = ρ(t )g, u = 0, + u T (κ T ) = γ. Viscosity η variations by four orders of magnitude in radial direction High main memory consumption due to the operators Operator evaluations of the Stokes equation are the performance critical parts of a (forward) simulation 8

9 Viscosity profiles (Fig. provided by M. Mohr) 9

10 Two-grid cycle (correction scheme) 10

11 Viscosity η variations by four orders of magnitude in radial direction: Adjust smoother Use matrix-dependent transfer operators Change coarse-grid operator High main memory consumption due to the operators: Matrix-free operator evaluations Operator evaluations of the Stokes equation are the performance critical parts of a (forward) simulation: Avoid indirect addressing by (semi-) structured meshes 11

12 Hierarchical Hybrid Grids 12

13 HHG - Combining finite element and multigrid methods FE mesh may be unstructured. What nodes to remove for coarsening? Not straightforward! Why not start from the coarse grid? The Hierarchical Hybrid Grids (HHG) concept Benjamin Bergen*: prototype Tobias Gradl: tuning, extensions and adaptivity * Dissertation in Erlangen, ISC award in Currently at Los Alamos Labs. 13

14 HHG mesh decomposition into primitives (2d-example) inner point ghost point volume edge memory representation geometric representation vertex 14

15 Regular refinement of a tetrahedral element 15

16 Smoothers 16

17 Operator: 15-point stencil in a refined tetrahedral element 17

18 Different Tetrahedral Shapes (a) Needle (b) Wedge (c) Spindle (d) Spade (e) Sliver (f) Cap Figure: Different classes of degenerated elements 18

19 Comparison between local Fourier analysis (LFA) and HHG Regular Needle Wedge Spindle Spade Sliver Cap LFA HHG edge ratio max angle min angle Table: LFA smoothing factors, µ ν1+ν2, and two grid convergence factors, ρ, of the four-color smoother for different tetrahedra Results of the LFA are provided by Francisco Gaspar 19

20 (a) Wedge (b) Spade (a) ν(1, 1) four-color Gauss-Seidel: ν(1, 1) line-wise smoothing: (b) ν(1, 1) four-color Gauss-Seidel: ν(1, 1) line-wise smoothing: ν(1, 0) zebra-plane smoothing:

21 (c) Needle (d) Spindle (c) ν(1, 1) four-color Gauss-Seidel: ν(1, 0) lex. plane smoothing: ν(1, 0) zebra-plane smoothing: (d) ν(1, 1) four-color Gauss-Seidel: ν(1, 0) zebra-plane smoothing:

22 (e) Sliver (f) Cap (e), (f)??? 22

23 (e) Sliver (f) Cap But: Acc. to Shewchuk: good tetrahedra are of two types: those that are not flat, and those that can grow arbitrarily flat without having a large planar or dihedral angle. The bad tetrahedra have error bounds that explode, and a dihedral angle or a planar angle that approaches 180 0, as they are flattened. J.R. Shewchuk. What Is a Good Linear Finite Element? - Interpolation, Conditioning, Anisotropy, and Quality Measures. In Proc. of the 11th International Meshing Roundtable (2002). 23

24 putting blocks together... (coarsest mesh is generated by Gmsh) 24

25 Locally adaptive smoothing We can predict the convergence rate for the single blocks of our domain by LFA. How to improve the convergence of the whole domain? 25

26 Locally adaptive smoothing We can predict the convergence rate for the single blocks of our domain by LFA. How to improve the convergence of the whole domain? Choose for each block: Different smoothers Optimal relaxation parameters Suitable number of pre-/post smoothing steps 25

27 Locally adaptive smoothing We can predict the convergence rate for the single blocks of our domain by LFA. How to improve the convergence of the whole domain? Choose for each block part: Different smoothers Optimal relaxation parameters Suitable number of pre-/post smoothing steps 26

28 Chaotic smoothing (mixed smoothers) Figure: Domain consisting of wedge, regular and needle tetrahedral types Shape edge ratio max angle min angle smoothing ν 1, ν 2 ρ Needle plane-wise 1, Regular color 2, Wedge line-wise 2, Table: Convergence factor for the single blocks Global convergence factor for the whole domain: using damping parameter at the interface:

29 Mantle convection Hierarchical Hybrid Grids Smoothers Geometric approximation Performance modeling Anisotropic smoothing (a) Connection in isotropic direction (b) Connection in anisotropic direction Gauss-Seidel: single region, line-wise smoother: (a) line-wise smoother: (b) line-wise smoother:

30 Locally adaptive smoothing We can predict the convergence rate for the single blocks of our domain by LFA. How to improve the convergence of the whole domain? Choose for each block: Different smoothers Optimal relaxation parameters Suitable number of pre-/post smoothing steps 29

31 Redistributing the number of smoothing steps Elements colored by convergence rates ν(pre-smoothing steps, post-smoothing steps) Keeping the total number of smoothing steps constant 30

32 Model Domain: Box in a Half Sphere Coarsest mesh: 293 elements 8 times refined unknowns lex. Gauss-Seidel β min = 0.241, β avg = Please note that this is a serial run! 31

33 Model Domain: Box in a Half Sphere We set up two conditions: 1. The required amount of communication is the same for all following methods. We are exchanging eight times all ghost boundary points at the interfaces during the smoothing procedure. 2. The total smoothing workload has to be comparable to the unoptimized version in the solving phase. The additional time for the setup phase strongly depends on the implementation of a LFA evaluation in some of the following smoothing strategies. 32

34 Local Adaptive Smoothing Smoothing strategy W-cycle V-cycle ω Unoptimized Exact block-wise solving Adaptive smoothing steps Additional interface smoothing Full damping Interior damping / Additional interface smoothing / 1.0 Adaptive interior damping (simplex) variable + Additional interface smoothing variable Combined Methods / 1.0 Table: Measured convergence factors for different smoothing strategies 33

35 Geometric approximation 34

36 Spherical refinement of an icosahedron (0) 35

37 Discretization with prismatic elements 36

38 Decomposition of prisms into three tetrahedra d e e a b b f f h c c g 37

39 Non-conforming decomposition of prisms into tetrahedra d e e a b b f f h c c g 38

40 Spherical refinement of an icosahedron (0) 39

41 Spherical refinement of an icosahedron (1) 40

42 Spherical refinement of an icosahedron (2) 41

43 Spherical refinement of an icosahedron (3) 42

44 Spherical refinement of an icosahedron (4) 43

45 Intersection through the spherical refinement - Each tetrahedral element ( 132k) was assigned to one Jugene compute core and further refined without curved boundaries. 44

46 Performance modeling 45

47 Storage of the variable viscosity field Since we store the viscosity like a variable, there are different ways to store them for linear finite elements: Element-wise More storage necessary for tetrahedral elements Clear location of a jump when dealing with jumping coefficients Node-wise Requires averaging No additional data layout necessary for structured tetrahedral elements 46

48 Viscosity averaging for a 15-point stencil e2 e3 e1 Coefficient averaging for each element Naive: 24 elements 3 = 72 adds By partial sums: 16 edges + 24 elements = 40 adds 47

49 Parallel generation of the coarsest mesh 1. Build up a local mesh on each process Loop over all elements in the mesh file: Read own macro-volume Loop over all elements in the mesh file: Read neighbouring macro-volumes 2. Map all other macro-primitives to processes Assign these primitives to the same process as the adjacent element with with the largest ID 3. Assign IDs to edge and face primitives Build tuples for all macro-edges and macro-faces (with one tuple entry per vertex ID) Sort the lists of tuples Assign IDs according to the index of the tuple in the sorted list 48

50 Mantle convection Hierarchical Hybrid Grids Smoothers Geometric approximation Performance modeling Weak and strong scaling on Jugene A triangle has a resolution of 6km2 ; hr = 3km in radial direction. 49

51 Seconds per V-cycle Terabytes Strong scaling on Jugene x1.93 x3.80 x1.85 x Number of compute cores Memory - constant viscosity Strong scaling - constant viscosity Memory - variable viscosity Strong scaling - variable viscosity 50

52 Performance modeling and analysis V-cycle 1. Computation: stencil evaluations 2. Network effects: bandwidth, latency, reductions including all multigrid levels and conjugate gradient on the coarsest mesh Assumptions one inner point on the coarsest mesh per processor hardware information exclusively extracted from IBM System Blue Gene Solution: Blue Gene/P Application Development 51

53 Time [s] BlueGene/P performance model (different setup) 0,35 0,30 0,25 0,20 0,15 0,10 0,05 0, #Cores Bandwidth CG Computation Measurement 52

54 Summary Introduction into challanges of mantle convection simulations Smoothing strategies for anisotropic elements and varying coefficients Presentation of a conforming tetrahedral decomposition of an icosahedral mesh Performance evalution and modeling Current work Multigrid implementation & evaluation of (jumping) variable coefficients for HHG Preparations for SuperMUC 53

55 Thank you for your attention! 54

56 Benchmarks on Blue Gene/P in Jülich (Jugene) Compute node: 4-way SMP processor Processortype: 32-bit PowerPC 450 core 850 MHz Cores: Overall peak performance: 1 Petaflops Main memory: 2 Gbytes per node (aggregate 144 TB) 55