/IMPL/SOLVER - Linear Solver Selection
Description
Selects linear solver.
Format
/IMPL/SOLVER/N
Data |
Description |
N |
Linear solver method number (solve Ax=b) = 0: set to 1 = 1: Iterative solver. Preconditioned Conjugate Gradient (PCG) = 2: Direct solver MUMPS = 3: Direct SMP: Boeing Solver (BCS) SPMD: Massively parallel multi-frontal solver (MUMPS) = 5: Mix - Two solvers will be used at same time. For instance, in SMP case, two solvers BCS and PCG will be used at same time. SMP: BCS and PCG (Iprec =5) SPMD: MUMPS and PCG (Iprec =5) = 7: Auto (Comment 4) Select automatic between solver direct and PCG (Iprec =5) |
Iprec |
Precondition methods = 0: set to 5 = 1: No preconditioner = 2: Diagonal Jacobi = 5: Factored approximate inverse |
It_max |
Maximum iteration number used for stop criteria = 0 or > NDOF: set to NDOF (system dimension) |
Itol |
Stop criteria for Preconditioned iterative solver flag = 0: set to 3 = 1: Relative Residual of original matrix (residual r =Ax-b, ||r|| < Tol * ||b||) = 2: Relative Residual of preconditioned matrix (||r|| < Tol * ||b’||) = 3: Relative Residual of preconditioned matrix (||r|| < Tol * ||b’|| * ||A’|| * ||x||) |
Tol |
Input tolerance for stop criteria = 0.0: set to 1.0e-5, if Itol =1 or 2 set to machine simple precision, if Itol =3 |
1. | When this keyword is not defined, all default values will be used. |
2. | Linear solvers are also used in nonlinear iterations. When /IMPL/NONLIN/N is used, the linear solver parameters, rather than default values can be defined in this keyword. |
3. | For Direct solvers (N=3), or Mix solver (N=5), mono-thread (nt=1) has been recommended for versions prior to 13.0; BCS is being used with mono-domain and MUMPS is used for multi-domain (np > 1). |
4. | N=5 (Mix): |
Two solvers will be used at same time. For instance, in SMP case, two solvers BCS and PCG will be used at same time.
N=7 (Auto):
PCG as the basic solver will be used first. It will factorize coefficient matrix (which is done by direct solver) and it is used as a pre-conditioner. The solver will be changed automatically to direct solver, if PCG performance is estimated slower than Direct solver. In this case, Direct solver will be used for the remainder of the run.
5. | Mix and Auto solvers are used for nonlinear implicit analysis. |
6. | PCG: solver (N=1) is fully Hybrid. Only this iterative solver is optimized for GPU. A GPU specific executable needs to be used. Currently, only Nvidia GPU is supported. |
See Also:
Implicit Finite Element Analysis
Compatibility Table of Implicit Solvers with Parallel Version