Overview

• Existing implementation
• Bins, Stencils, Pairs
• Multi-level implementation
• Bins, Stencils, Pairs, ...
• Communication
• Results
• Polydisperse
• Bidisperse
• Communication
• Summary / observations

LAMMPS neighbour bins

Start with cell list! [Alder and Wainwright (1959)]

Implementation: NBinStandard extends NBin

    "neighbor 0.5 bin"
    "neighbor 0.5 nsq"
    

LAMMPS neighbour stencils (1)

Small and larger particle [t'Veld et al.(2008)]

Implementation: various stencils extend NStencil

    "neighbor 0.5 multi"
    "newton off"
    

LAMMPS neighbour pair lists

[Verlet (1967)]

Implementation: various classes extend NPair

• For each particle, maintain list a of neighbours
• List is searched to compute actual interactions

Final /bin/stencil/pair selection is reported...

  "(1) pair gran/hooke/history/multi, perpetual
      attributes: half, newton on, size, history
      pair build: half/size/bin/newton
      stencil: half/bin/3d/newton
      bin: standard"
    

LAMMPS stencils again

Problem

Large separation of scales not handled well by stencil

Multi-level implementation

Bin different size particles in different size cell lists!

Particles identified by standard type attribute

Implementation: NBinByType extends NBin

    "neighbor 0.5 bytype"
    

Multi-level implementation: Stencils

Hierarchy favours compact stencils:

• Small-small neighbours located in small cell list
• Large-large neighbours located in large cell list

What about cross-type interactions?

Implementation: new stencils extend NStencil

Communication

Domain decomposition plus message passing

• Multi approach: different sizes have different cut offs
• Small is $r_{sl}$; large is $r_{ll}$


• New approach:
• Small is $r_{ss}$; large still $r_{ll}$
• ("Newton" only)

Polydisperse result

Model system based on sand

• Real distribution scaled to set $d_{max}/d_{min}$
• We use two types (two levels small/large)
• How do we split the size distribution into two types?
  

Bidisperse result

Exactly two sizes

• $R = 1$ and $R = R$ for $R = 10, 20, 40, 80, 100$

Communication / scaling

    "comm_modify mode multi cutoff/bytype"
    

Summary / observations

• Existing multi approach extended to multi-level picture
• Effective for polydisperse systems
• Good for bidisperse systems to 100:1


• More robust communication?
• Decouple from type system?

Acknowledgements

Ishan Srivastava of Sandia for making available some extra LAMMPS granular classes.

Additional slides...