SOLVER ADAPTATION TO A SLAE FLUX
V. A. Erzunov, Yu. G. Bartenev
VANT. Ser.: Mat. Mod. Fiz. Proc 2021. Вып.1. С. 68-79.
At RFNC-VNIIEF, a number of codes for the computational physics problems use the PMLP/ParSol library for iteratively solving linear systems of algebraic equations with sparse matrices, both in serial and parallel compute modes. SLAEs are solved on the distributed-memory (message passing interface, MPI) and shared-memory (OpenMP interface) platforms by calling for an appropriate solver with the given preconditioner. To reduce the SLAE solution time, it is useful to have a capability of changing over from one solver to another during the compute process that allows the user not to perform initial selection of an optimum solver for the problem and replace it by another solver, which is most appropriate for the SLAE properties changed within the problem runtime. This means it is necessary to have an adaptive mechanism for automatically selecting solvers for the problem and the particular stage of simulating the physical process. The adaptive mechanism parameters are set in the PMLP/ParSol library file, where the methods tested and their parameters are specified for the optimum SLAE solution process. After the SLAE solution using one method has been examined, the adaptive mechanism allows testing another method and continues computations using the most optimum one of all the tested methods. The adaptive mechanism of selecting a solver has been verified on nonlinear and linear heat conduction problems using various RFNC-VNIIEF codes. Computations demonstrate that the adaptive mechanism reduces the SLAE solution time and the entire problem runtime. Keywords: system of linear algebraic equations (SLAE), sparse matrices, distributed-memory computing systems, multicore shared-memory processors, adaptive mechanism.
|
English
