4/30/04LSU 2004 Cactus Retreat1 Toward Relativistic Hydrodynamics on Adaptive Meshes Joel E. Tohline Louisiana State University tohline.

4/30/04 LSU 2004 Cactus Retreat 1

Toward Relativistic Hydrodynamics on Adaptive Meshes

Joel E. Tohline

Louisiana State University

http://www.phys.lsu.edu/~tohline

4/30/04 LSU 2004 Cactus Retreat 2

Principal Collaborators• Simulations to be shown today:

– Shangli Ou – (LSU)– Mario D’Souza – (LSU)– Michele Vallisneri – (Caltech/JPL)

• Code development over the years:– John Woodward – (Valtech; Dallas, Texas)– John Cazes – (Stennis Space Center; Stennis, Mississippi)– Patrick Motl – (Colorado)

• Science:– Juhan Frank (LSU)– Lee Lindblom (Caltech)– Luis Lehner (LSU)– Jorge Pullin (LSU)

4/30/04 LSU 2004 Cactus Retreat 3

Show 3 Movies

• Nonlinear development of the r-mode in young neutron stars [w/ Lindblom & Vallisneri] http://www.cacr.caltech.edu/projects/hydrligo/rmode.html

• Nonlinear development of the secular bar-mode instability in rapidly rotating neutron stars [w/ Ou & Lindblom] http://paris.phys.lsu.edu/~ou/movie/fmode/new/fmode.b181.om4.2e5.mov

• Mass-transferring binary star systems [w/ D’Souza, Motl, & Frank] http://paris.phys.lsu.edu/~mario/models/q_0.409_no_drag_3.8orbs/movies/q_0.409_no_drag_3.8orbs_top.mov

4/30/04 LSU 2004 Cactus Retreat 4

Storyline• Present Algorithm – has been producing publishable astrophysical

results for over 20 years: – Entirely home-grown code outside of Cactus environment– Manual domain decomposition– Explicit message-passing using mpi– Visualizations on serial machines (generally, post-processing)

• Plans for this calendar year:– Move present algorithm into Cactus environment

• Over the next few years, modify algorithm to:– Follow relativistic hydrodynamical flows on adaptive mesh– Accept evolving space-time metric– Visualize results “in parallel” with dynamical evolution

4/30/04 LSU 2004 Cactus Retreat 5

Storyline• Present Algorithm – has been producing publishable astrophysical

results for over 20 years: – Entirely home-grown code outside of Cactus environment– Manual domain decomposition– Explicit message-passing using mpi– Visualizations on serial machines (generally, post-processing)

• Plans for this calendar year:– Move present algorithm into Cactus environment

• Over the next few years, modify algorithm to:– Follow relativistic hydrodynamical flows on adaptive mesh– Accept evolving space-time metric– Visualize results “in parallel” with dynamical evolution

4/30/04 LSU 2004 Cactus Retreat 6

Storyline• Present Algorithm – has been producing publishable astrophysical

results for over 20 years: – Entirely home-grown code outside of Cactus environment– Manual domain decomposition– Explicit message-passing using mpi– Visualizations on serial machines (generally, post-processing)

• Plans for this calendar year:– Move present algorithm into Cactus environment

• Over the next few years, modify algorithm to:– Follow relativistic hydrodynamical flows on adaptive mesh– Accept evolving space-time metric– Visualize results “in parallel” with dynamical evolution

4/30/04 LSU 2004 Cactus Retreat 7

Present Algorithm

• Select grid structure and resolution• Construct initial configuration• Perform domain decomposition• While t < tstop

– Determine Newtonian gravitational accelerations– Push fluid around on the grid using Newtonian dynamics– If mod[ t , (orbital period/80) ] = 0

• Dump 3-D dataset for later visualization

– EndIf

• EndWhile• Visualize results

4/30/04 LSU 2004 Cactus Retreat 8

Principal Governing Equations

4/30/04 LSU 2004 Cactus Retreat 9

Principal Governing Equations

4/30/04 LSU 2004 Cactus Retreat 10

Present Algorithm

• Select grid structure and resolution• Construct initial configuration• Perform domain decomposition• While t < tstop

– Determine Newtonian gravitational accelerations– Push fluid around on the grid using Newtonian dynamics– If mod[ t , (orbital period/80) ] = 0

• Dump 3-D dataset for later visualization

– EndIf

• EndWhile• Visualize results

4/30/04 LSU 2004 Cactus Retreat 11

Present Algorithm

• Select grid structure and resolution• Construct initial configuration• Perform domain decomposition• While t < tstop

– Determine Newtonian gravitational accelerations– Push fluid around on the grid using Newtonian dynamics– If mod[ t , (orbital period/80) ] = 0

• Dump 3-D dataset for later visualization

– EndIf

• EndWhile• Visualize results

4/30/04 LSU 2004 Cactus Retreat 12

Present Algorithm

• Select grid structure and resolution• Construct initial configuration• Perform domain decomposition• While t < tstop

– Determine Newtonian gravitational accelerations– Push fluid around on the grid using Newtonian dynamics– If mod[ t , (orbital period/80) ] = 0

• Dump 3-D dataset for later visualization

– EndIf

• EndWhile• Visualize results

Serial

Serial

4/30/04 LSU 2004 Cactus Retreat 13

Present Algorithm

• Select grid structure and resolution• Construct initial configuration• Perform domain decomposition• While t < tstop

– Determine Newtonian gravitational accelerations– Push fluid around on the grid using Newtonian dynamics– If mod[ t , (orbital period/80) ] = 0

• Dump 3-D dataset for later visualization

– EndIf

• EndWhile• Visualize results

Parallel

4/30/04 LSU 2004 Cactus Retreat 14

Parallel Code’s Chronological Evolution

• John Woodward:– 8,192-processor MasPar @ LSU

• John Cazes:– CM5 @ NCSA; T3D/E @ SDSC

• Patrick Motl:– mpi on T3E @ SDSC; SP2/3 @ SDSC & LSU

• Michele Vallisneri:– HP Exemplar @ CACR

• Mario D’Souza & Shangli Ou:– SuperMike (1,024-proc Linux cluster) @ LSU

• Shangli Ou:– Tungsten (2,560-proc Linux cluster) @ NCSA

Early 90’s

Mid-90’s

Late 90’s

2000

2002-03

2004

4/30/04 LSU 2004 Cactus Retreat 15

Select Grid Structure and Resolution

• Unigrid, cylindrical mesh

• Fixed in time

• Typical resolution– Single star: 66 x 128 x 130 (as

shown on the left)

– Binary system: 192 x 256 x 98

4/30/04 LSU 2004 Cactus Retreat 16

Select Grid Structure and Resolution

4/30/04 LSU 2004 Cactus Retreat 17

Need for Non-unigrid and Adaptive Meshes

4/30/04 LSU 2004 Cactus Retreat 18

Perform Domain Decomposition

• Grid resolution 192 x 256 x 96• 64 processors• Distribute 192 x 96 (R,Z) grid

across 8 x 8 processor array• Leave angular zones “stacked”

in memory• Result: Each processor has

data arrays of size 24 x 256 x 12

• I/O: Scramble and unscramble handled manually

Z

R

4/30/04 LSU 2004 Cactus Retreat 19

Determine Newtonian Gravitational Accelerations(Three-dimensional Elliptic PDE on cylindrical mesh)

4/30/04 LSU 2004 Cactus Retreat 20

Principal Governing Equations

4/30/04 LSU 2004 Cactus Retreat 21

Determine Newtonian Gravitational Accelerations(Three-dimensional Elliptic PDE on cylindrical mesh)

• Perform FFT (in memory) in azimuthal coordinate direction reduce to decoupled set of (256) two-dimensional Helmholtz equations.

• Use ADI (alternating direction implicit) to solve each 2-D equation:– Data transpose– 1-D, in-memory ADI sweep– Data transpose– 1-D, in-memory ADI sweep– Data transpose– Etc.

• Inverse FFT

Z

R

4/30/04 LSU 2004 Cactus Retreat 22

Determine Newtonian Gravitational Accelerations(Three-dimensional Elliptic PDE on cylindrical mesh)

• Perform FFT (in memory) in azimuthal coordinate direction reduce to decoupled set of (256) two-dimensional Helmholtz equations.

• Use ADI (alternating direction implicit) to solve each 2-D equation:– Data transpose– 1-D, in-memory ADI sweep– Data transpose– 1-D, in-memory ADI sweep– Data transpose– Etc.

• Inverse FFT

Z

m

4/30/04 LSU 2004 Cactus Retreat 23

Determine Newtonian Gravitational Accelerations(Three-dimensional Elliptic PDE on cylindrical mesh)

• Perform FFT (in memory) in azimuthal coordinate direction reduce to decoupled set of (256) two-dimensional Helmholtz equations.

• Use ADI (alternating direction implicit) to solve each 2-D equation:– Data transpose– 1-D, in-memory ADI sweep– Data transpose– 1-D, in-memory ADI sweep– Data transpose– Etc.

• Inverse FFT

R

m

4/30/04 LSU 2004 Cactus Retreat 24

Visualize Results

• Specify isodensity surface(s)

• Find vertices and polygons on each surface (using marching cubes algorithm)

• Write out vertices & polygons in “OBJ” format

• Delete 3-D dataset

• Utilize “Maya” to render nested surfaces (from pre-specified viewer orientation)

• Write out TIFF image (typically 640 x 480)

• Generate .mov

4/30/04 LSU 2004 Cactus Retreat 25

Future Algorithm

• Select grid structure and resolution and [preferred AMR thorn]• [We] Construct initial configuration• [Let Cactus] Perform domain decomposition• While t < tstop

– [Call GR Group’s thorn] Determine structure of space-time metric– [We (or Whisky thorn)] Push fluid around on the grid using Relativistic dynamics– If mod[ t , (orbital period/80) ] = 0

• Generate vertices and polygons in parallel• Spawn “Maya” rendering task on additional processor(s)

– EndIf– [Call AMR thorn] Modify mesh, as necessary

• EndWhile• [Use Cactus thorn] Write article and Publish results

4/30/04 LSU 2004 Cactus Retreat 26

Future Algorithm

• Select grid structure and resolution and [preferred AMR thorn]• [We] Construct initial configuration• [Let Cactus] Perform domain decomposition• While t < tstop

– [Call GR Group’s thorn] Determine structure of space-time metric– [We (or Whisky thorn)] Push fluid around on the grid using Relativistic dynamics– If mod[ t , (orbital period/80) ] = 0

• Generate vertices and polygons in parallel• Spawn “Maya” rendering task on additional processor(s)

– EndIf– [Call AMR thorn] Modify mesh, as necessary

• EndWhile• [Use Cactus thorn] Write article and Publish results

4/30/04 LSU 2004 Cactus Retreat 27

THE END