Particle Method Based Applications and their Parallel Optimizations towards Exascale.

Date: August 11
Time: 16:00--18:00
Room: 406

(Note: Click title to show the abstract.)

Guo, Xiaohu (Sci. & Tech. Facilities Council)

Abstract: A large number of industrial physical problems can be modeled using particle-based methods. Particle descriptions can be used for the simulation of continuum systems as in the case of discrete fluid or solid elements in smooth particle hydrodynamics (SPH) and vorticity-carrying fluid elements in vortex methods (VM); or for inherently discrete systems as in gravitational particles for astrophysics, dissipative particle dynamics (DPD) for mesoscale polymer descriptions, atomistic molecular dynamics (MD) simulations, and charged particles in plasma physics(PIC).
The dynamics of particle methods are governed by the interactions of the N computational particles resulting in an N-body problem with a computational cost that scales nominally as O(N2). For short-ranged particle interactions, the computational cost scales linearly with the number of particles. Therefore, the parallel implementations of these methods are critical for application purpose.
In recent published DOE exascale math report, particles based methods have been identified as ˇ°well-suited for exascale computingˇ±. The reason is that particles based methods provide extremely fine-grained parallelism and each particle can be compute independently and allow the exploitation of asynchrony. Particles model can effectively use single precision accuracy due to the associated statistical noise. And their statistical nature also makes the application software resilient to both soft and hard faults. Methods in this category include Monte Carlo, smoothed particle hydrodynamics, and particle-in-cell techniques.
Efficient parallel particle method based application libraries typically require the following efficient kernel implementation: domain decomposition, dynamic load balancing, optimized data mapping (structured and unstructured communication), parallel file I/O, nearest neighbor lists searching routines for building trees, particle-to-mesh, and mesh-to-particle interpolation, sparse linear solver for incompressible problems.
The present proposal aims to identify the common efficient implementation of the above kernels and their high performance optimizations among above different applications for the future exascale systems. In the mean time, the following implementation factors will be considered:
• Exploiting the symmetry of the particle interactions requires sending back of ghost contributions to the proper real particle
• The simultaneous presence of particles and meshes prohibits a single optimal way of parallelization
• Complex-shaped computational domains and strong particle inhomogeneities require spatially adaptive 
domain decompositions
• Particle motion may invalidate the existing domain decomposition causing rising load imbalance, and com
plicate the implementation of multi-stage integration schemes
• Inter-particle relations constrain decompositions and data assignment
• Efficient implementations on many core architectures.

Highly Scalable Parallel Toolkit for Solving a complex, highly nonlinear and distorted flow using Incompressbile Smoothed Particle Hydrodynamics
Guo, Xiaohu (Sci. & Tech. Facilities Council)

Feature-scale simulations of particulate slurry flows in chemical mechanical polishing by SPH
Shao, Sihong (Peking Univ.)
Yan, Changhao (Fudan Univ.)
Cai, Wei (Univ. of North Carolina at Charlotte)

Incompressible Smoothed Particle Hydrodynamics
Lind, Steven (Univ. of Manchester)

Variational Symplectic Algorithm for Kinetic Plasma Simulation
Xiao, Jianyuan (Univ. of Sci. & Tech. of China)


Code: Type-Date-Time-Room No.
Type : IL=Invited Lecture, SL=Special Lectures, MS=Minisymposia, IM=Industrial Minisymposia, CP=Contributed Papers, PP=Posters
Date: Mo=Monday, Tu=Tuesday, We=Wednesday, Th=Thursday, Fr=Friday
Time : A=8:30-9:30, B=10:00-11:00, C=11:10-12:10, BC=10:00-12:10, D=13:30-15:30, E=16:00-18:00, F=19:00-20:00, G=12:10-13:30, H=15:30-16:00
Room No.: TBA