Octree Methods for Multiscale Discretization
Abstract
This project will investigate various discretization frameworks for forests (collections) of Octree Cartesian grids for eventual implementation on modern HPC architectures. This work will also address the related question of effective boundary conditions that will make possible advances in multiscale modeling through coupling between molecular dynamics simulations and continuum-level dynamics. The main challenge when considering parallel computing is to design algorithms that maximize the amount of computation local to each processor, while minimizing communication. The first important task is therefore to effectively partition the computational grid, i.e. to evenly distribute the entire grid among the available processors to ensure a balanced workload. This will be done through orderings to map 2D and 3D arrays of cells into a linear ordering, and through development of local octree data structures with which to discretize differential operators in such a manner as to closely or exactly match leaves to local computational cells. Another effort will be pursued to construct unconditionally stable methods that can take large time steps. However, the challenge with taking large time steps in a parallel environment is that this requires fetching data not necessarily located locally, increasing the amount of communication. This will require new methods that enable arbitrarily large time steps yet still permit scalable performance .
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 12, 2017
- Source ID
- W911NF1610136
Entities
People
- Frédéric Gibou
Organizations
- Army Contracting Command
- United States Army
- University of California, Santa Barbara