Parallel Multilevel Decomposition Methods for Fluid Plasma Models
Abstract
1. Multilevel Decomposition. Full implementations of both domain and range versions were developed and the comparative advantages and disadvantages were studied and documented. Numerical tests and model analyses showed these methods to be effective for current and foreseeable large-scale parallel architectures. 2. Adaptive AMG. New adaptive versions of AMG/SA were developed and tested, showing that the large-scale performance characteristics of these full-system methods to be superior to conventional approaches. 3. Jacobian-Free Newton-Krylov Methods. New physics-based preconditioners for the two-fluid problems were developed that exploit the new, ultra-scalable multilevel algorithms. New strategies for reducing preconditioner setup costs during the JFNK iterations were developed and implemented. 4. Space-Time Parallelization. A new multilevel algorithm was developed for implicit time-stepping schemes that coarsens in space and time and its efficacy for space-time grids was demonstrated. 5. Fault Resilience. A software layer for fault detection and recovery was developed that seamlessly integrates into the multilevel decomposition methods. Efficiency of the proposed fault detection and recovery method was carefully demonstrated.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 26, 2016
- Accession Number
- AD1005446
Entities
People
- Kristian R. Beckwith
- Luke N. Olson
- Stephen F. Mccormick