Particle Simulations with HOOMD-blue Joshua A. Anderson S6256 - PowerPoint PPT Presentation

Particle Simulations with HOOMD-blue Joshua A. Anderson S6256 NVIDIA GTC April 7, 2016 9:00-9:50am T HE G LOTZER G ROUP

HOOMD-blue General purpose particle simulation toolkit Molecular dynamics, hard particle Monte Carlo Executes simulations fast on GPUs (also runs on CPUs) Open source >127 peer-reviewed articles 2 T HE G LOTZER G ROUP

History 2007-02 - CUDA public beta 2007-03 - APS March meeting 2007-09 - Paper submitted 2008-02 - v0.6 : C++ benchmarks only, lj and harmonic bonds only 2008-08 - v0.7: Python job scripts 2008-12 - v0.8: First external contribution - FENE bonds, Langevin dynamics 3 T HE G LOTZER G ROUP

History 2015-09 - v1.2: Bounding volume hierarchy neighbor list 2015-12 - v1.3: Anisotropic particle integrators, wall potentials, load balancing 2016-?? - v2.0: Hard particle Monte Carlo 4 T HE G LOTZER G ROUP

Research using HOOMD Microspheres Shape allophiles Quasicrystal formation Active particles Digital alchemy Hexatic phases 5 T HE G LOTZER G ROUP

Molecular dynamics v i ( t ) r i ( t ) ~ ~ q i ( t ) ~ ! i ( t ) ~ Compute interactions r i ( t + � t ) v i ( t + � t ) ~ ~ q i ( t + � t ) ! i ( t + � t ) ~ ~ 6 T HE G LOTZER G ROUP

Hard particle Monte Carlo (HPMC) q i ( A ) ~ r i ( A ) ~ Random trial move Accept or reject r i ( B ) q i ( B ) ~ ~ Joshua A Anderson, M Eric Irrgang, Sharon C Glotzer (2016) Scalable Metropolis Monte Carlo for simulation of hard shapes, Computer Physics Communications 7 T HE G LOTZER G ROUP

HPMC Shape overlap checks + See GTC 2014 talk 8 T HE G LOTZER G ROUP

HPMC block level queue (a) Initialization Trial move Circumsphere check Overlap check Overlap divergence Early exit divergence Thread Time (0.5 ms total) (b) Time 9 T HE G LOTZER G ROUP

HPMC on P100 10 T HE G LOTZER G ROUP

GPU strategy in HOOMD Perform all computations on the GPU (see GTC 2010-2013 talks) Keep data on the GPU all the time Run multiple threads per particle and warp-level reduction/scan Auto-tune launch parameters during run time (see GTC 2015 talk) 11 T HE G LOTZER G ROUP

Architecture User's python script HOOMD python library matplotlib scipy mpi4py ... HOOMD C++ library CUDA MPI Hardware 12 T HE G LOTZER G ROUP

Installing binaries with conda http:// conda.pydata.org/miniconda.html 13 T HE G LOTZER G ROUP

Compile from source Requirements: C++ compiler Python Numpy Boost C++ library CMake Optional: CUDA MPI git sphinx 14 T HE G LOTZER G ROUP

Compile from source 15 T HE G LOTZER G ROUP

Example: DPD polymers import hoomd from hoomd import md hoomd.context.initialize(); hoomd.init.read_gsd(filename='polymers.gsd') nl = hoomd.md.nlist.cell() dpd = md.pair.dpd(r_cut=1.0, T=1.0, nlist=nl) nl.reset_exclusions() dpd.pair_coeff.set('A', 'A', A=25.0, gamma = 4.5) dpd.pair_coeff.set('A', 'B', A=80.0, gamma = 4.5) dpd.pair_coeff.set('B', 'B', A=25.0, gamma = 4.5) harmonic = hoomd.md.bond.harmonic() harmonic.bond_coeff.set('polymer', k=330.0, r0=0.84) md.integrate.mode_standard(dt=0.02) md.integrate.nve(group=hoomd.group.all()) hoomd.run(10000) 16 T HE G LOTZER G ROUP

Running the example 17 T HE G LOTZER G ROUP

T HE G LOTZER G ROUP

Example: Active matter import hoomd, numpy from hoomd import md hoomd.context.initialize() hoomd.init.read_gsd('active.gsd') all = hoomd.group.all() N = len(all) nl = hoomd.md.nlist.cell() lj = md.pair.lj(r_cut=1.0, nlist=nl) lj.pair_coeff.set('A', 'A', epsilon=1.0, sigma=1.0/2**(1.0/6.0)) activity = [ (((numpy.random.rand(3) - 0.5) * 2.0)) for i in range(N)] for i in range(N): activity[i][2] = 0; activity[i] = tuple(activity[i]) active_force = md.force.active(group=all, seed=123, f_lst=activity, rotation_diff=0.005, orientation_link= False ) md.integrate.mode_standard(dt=0.001) bd = md.integrate.brownian(group=all, T=0.0, seed=123, dscale=1.0) hoomd.run(100000) 19 T HE G LOTZER G ROUP

20 T HE G LOTZER G ROUP

BVH tree overview BVH trees are common in ray tracing HOOMD uses them for neighbor searches (optional) Generate and query on the GPU M. P. Howard, J. A. Anderson, A. Nikoubashman, S. C. Glotzer, and A. Z. Panagiotopoulos, Comput. Phys. Commun. , Mar. 2016. 21 T HE G LOTZER G ROUP

Example: Large and small particles import hoomd, numpy from hoomd import md hoomd.context.initialize() hoomd.init.read_gsd('bigsmall.gsd') nl = md.nlist.tree() lj = md.pair.slj(r_cut=2**(1.0/6.0), nlist=nl) lj.pair_coeff.set('A', 'A', epsilon=1.0, sigma=1.0) md.integrate.mode_standard(dt=0.005) lvn = md.integrate.langevin(group=hoomd.group.all(), T=1.0, seed=123) lvn.set_gamma('A', 1.0) hoomd.run(3000000) 22 T HE G LOTZER G ROUP

23 T HE G LOTZER G ROUP

Example: Hard Hexagons import hoomd from hoomd import hpmc hoomd.context.initialize(); hoomd.init.read_gsd(filename='hexagons.gsd') hexagon = [[0.5,0], [0.25,0.433012701892219], [-0.25,0.433012701892219], [-0.5,0], [-0.25,-0.433012701892219], [0.25,-0.433012701892219]]; mc = hpmc.integrate.convex_polygon(seed=123, d=0.15, a=1); mc.shape_param.set('A', vertices=hexagon); hoomd.run(200) 24 T HE G LOTZER G ROUP

25 T HE G LOTZER G ROUP

Example: Jupyter notebook 26 T HE G LOTZER G ROUP

Additional features Integration: NVE, NVT, NPT, NPH, langevin, brownian, • Berendsen, DPD, FIRE minimization Orientational degrees of freedom • Pair potentials: CGCMM, DPD, LJ, Gaussian, Mie, • Moliere, Morse, Yukawa, ZBL, table Bond: FENE, Harmonic, table, OPLS • Anisotropic potentials: Gay-berne, dipole • Wall potentials • PPPM electrostatics • HPMC shapes: convex polygon, simple polygon, • convex spheropolygon, convex polyhedron, convex spheropolyhedron, general polyhedron, ellipsoid, differences of spheres, unions of spheres, faceted spheres 27 T HE G LOTZER G ROUP

Software engineering Object oriented design Template functors Code review / pull requests Unit testing Validation testing Conda recipes Read the docs 28 T HE G LOTZER G ROUP

T HE G LOTZER G ROUP

T HE G LOTZER G ROUP

Read The Docs 31 T HE G LOTZER G ROUP

Acknowledgements HOOMD-blue + HPMC: codeblue.umich.edu/hoomd-blue/ 2.0 will be available soon..... Thanks to all HOOMD-blue users and contributors! Research supported by the National Science Foundation, Division of Materials Research Award # DMR 1409620. 32 T HE G LOTZER G ROUP