Simplifying Partitioning Complexities by Using a Common Data Hub
Abstract
Most scalable, high-performance simulations of interest to the Department of Defense involve calculations on a geometric structure. These topologies (unconnected points, unstructured meshes, rectilinear grids, etc.) are then partitioned across many processors. The simulation proceeds in parallel with information being communicated between processors as necessary. If the topology of the mesh changes (crack propagation, adaptive mesh refinement, etc.), the mesh must be repartitioned before efficient computation may continue. In addition, if data must be shared between various topologies (Coupled Eulerian-Lagrangian simulation), the complexities of transferring data between meshes, in parallel, can be significant. The Interdisciplinary Computing Environment has defined a common data model and format that can efficiently consolidate large quantities of computed data during runtime. Initially used for runtime visualization of parallel simulations, this data hub is now being applied to the computational simulations themselves in order to alleviate the complexities of communicating data between different topologies and the repartitioning of evolving topologies. Data can be written to the data hub using one partitioning scheme and read back with a totally different scheme. In the situation of different topologies, programs need only be cognizant of their partitioning and the "overall" topology of the opposing mesh.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 2002
- Accession Number
- ADA406684
Entities
People
- Jerry A. Clarke
- Raju R. Namburu
Organizations
- United States Army Research Laboratory