[petsc-users] performance issue solving multiple linear systems of the same size with the different preconditioning methods

[Jed Brown]

2011/8/27 Алексей Рязанов <ram at ibrae.ac.ru>

> So. All memory and time issues appeared because of my KSPConvergedTest
> function, which calls KSPBuildResidual to monitor true
> residual. KSPBuildResidual incorrectly works with memory. As soon as I've
> commented the line with KSPBuildResidual, all troubles gone.  So now I use
> default monitor and have to monitor preconditioned residual. This is
> inconvenient for me, but I cant figure out, how can I set up using true
> monitor for all methods without any memory or performance leaks. I'll try to
> get used t deal with preconditioned norm.


1. You can actually run the Krylov method with unpreconditioned residuals by
using -ksp_right_pc (some output is inconsistent with 3.1 this way, it's
fixed in petsc-dev and you can just use -ksp_norm_type unpreconditioned).
This is inexpensive and does the right thing.

2. You can run the Krylov method in the preconditioned norm, but monitor the
true residual using -ksp_monitor_true_residual. The code is in
src/ksp/ksp/interface/iterativ.c KSPMonitorTrueResidualNorm(). You can also
use this code as a template to make a similar custom monitor without memory
leaks. This calls KSPBuildResidual() which is very expensive for GMRES, but
is cheap for some other methods (including GCR which is closely related).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20110827/4c1752d4/attachment-0001.htm>