BCGSL – Biconjugate gradient stabilized variation

Usage:

-ksp_type bcgsl

Options:

-ksp_bcgsl_ell <ell>

Number of Krylov search directions

-ksp_bcgsl_cxpol

Use a convex function of the MR and OR polynomials after the BiCG step

-ksp_bcgsl_xres <res>

Threshold used to decide when to refresh computed residuals

A slight variant of the Enhanced BiCGStab(L) algorithm. The variation concerns cases when either kappa02 or kappa12 is negative due to round-off. Kappa0 has also been pulled out of the denominator in the formula for .

For about the same computational load, one application of BCGSL is similar to L steps of BCGS.
In one step of BCGSL, GMRES(L) is applied once while in L steps of BCGS, GMRES is applied L times.
One step of GMRES(L) can be expected to give super-linear convergence compared to regular GMRES.

ADH Benefits:

Improves Het_Column8 slightly when combined with preconditioner ICC.
Improves Backstep speed dramatically.

Hints: Generally L=2 or L=4 is optimal.

Back to Main Page