Automated Solver Selection for Nuclear Engineering Simulations
- RNET Technologies, Inc.
- RNET Technologies
- University of Oregon
Large scale numeric simulation codes often rely on linear solvers as a basic building block. These linear solvers tend to be the most compute resource intensive code block in most of these simulations, and thus also consume a major portion of large supercomputer resources. The linear solver choice can significantly affect the portability numeric simulations tools to a range of computing platforms and thus to a wider nuclear industry. This choice further needs to be made in a dynamic context such as a time evolving system. While the experience and expertise provided by the domain scientists to artfully choose the best solver cannot be underestimated, it is desirable to have toolkit-enabled functionality that automatically chooses the solver that is both appropriate for the computational problem at hand and also is efficient on the given hardware. Automatic selection of an appropriate solver can lead to benefits such as reduced memory requirement, lower execution time, fewer synchronization points in a parallel computation and so forth. However, there is no governing theory for finding the best numerical subroutines, or the theory is too expensive to compute. There is no uniformly best method that is independent of the characteristics of the linear system, even for a simplified computing environment. The choice of the optimal method is in practice determined by experimentation and numerical folklore. This project builds on research efforts toward using cheap matrix features and machine learning to accurately predict the optimal linear solver configuration for a given linear system. This project includes identifying and extracting relevant features from the linear system, and specifying selection criteria for algorithmic choices. The automatically chosen solvers exhibited an order of magnitude speedup over the default solvers for sparse linear systems drawn from CFD and magneto thermodynamics applications. The software framework allows these novel machine learning approaches to be used seamlessly inside advanced numerical tools. The primary product developed in this SBIR is the SolverSelector library. This library enables automatic solver selection in numerical applications. The library ships with support for automatic solver selection in PETSc and can be extended to fit other linear algebra packages and applications. Some of the features of the SolverSelector library include: binary solver classification with prediction based solver ranking, binary solver classification with efficient top down solver ranking, non-binary solver classification for automatic linear solver selection in numerical simulations, PETSc interface for automatic krylov solver and preconditioner selection in PETSc applications, and Parallel scaling through discrete sampling and linear interpolation.
- Research Organization:
- RNET Technologies
- Sponsoring Organization:
- USDOE Office of Nuclear Energy (NE)
- DOE Contract Number:
- SC0013869
- OSTI ID:
- 1722955
- Type / Phase:
- SBIR (Phase II)
- Report Number(s):
- RNET-13869-1
- Country of Publication:
- United States
- Language:
- English
Similar Records
PETSc Users Manual (Rev. 3.4)
PETSc Users Manual (Rev. 3.5)