[petsc-users] Speeding-up petsc

[Bruno Ramon Batista Fernande]

Hello guys,

I am using petsc is a solver for a finite difference PDE solver.

Basically the coefficient matrix (say A in Ax = b) is well conditioned,
symmetric, and up to 7 diagonals. Also, no pivoting is necessary for such
problem and it is always diagonally dominant.

I am using all default options of petsc but I still feel like I could get
better results by choosing different answers. By default I mean: GMRES
solver preconditioned on left with ILU(0) and a classical Gram-Schmidt to
orthogonalize against krylov vector.

I am using MAT_STRUCTURALLY_SYMMETRIC option and
atol = 1.0e-30
dtol = 1.0e5

You guys have any idea of how to further improve performance of the solver?

Regards,
Bruno

[Jed Brown]

Post by Bruno Ramon Batista Fernande
Hello guys,
I am using petsc is a solver for a finite difference PDE solver.
Basically the coefficient matrix (say A in Ax = b) is well conditioned,
symmetric, and up to 7 diagonals. Also, no pivoting is necessary for such
problem and it is always diagonally dominant.
I am using all default options of petsc but I still feel like I could get
better results by choosing different answers. By default I mean: GMRES
solver preconditioned on left with ILU(0) and a classical Gram-Schmidt to
orthogonalize against krylov vector.

You have told us nothing about the performance of this solver. Always
include -log_summary -ksp_view.

Post by Bruno Ramon Batista Fernande
I am using MAT_STRUCTURALLY_SYMMETRIC option and
atol = 1.0e-30
dtol = 1.0e5

Why these options?

How accurate of a solution do you need?