[petsc-users] EPS iteration number

[Xujun Zhao]

Hi Jose,

This is only a very simple test. 2 outer iterations perform 24 MatMult
operation, which means I have to solve a Finite element system with
millions of DOFs for 24 times. This is only for a polymer system with two
beads. For a polymer chain with hundreds of beads and springs, I need more
outer iterations to get its max/min eigenvalues, which means I have to
perform more computations of the finite element system. Those solves happen
only for one time step. When polymer beads move, I need to redo everything.

In other words, the eigenvalues (max and min) have to be solved every time
step or every several time steps. Therefore, it is very expensive! My
question is that if It is possible to use the eigenvalues obtained from
last time step as an initial guess to accelerate the convergence. Thank you
for your help.

Xujun





On Wed, Sep 16, 2015 at 4:18 PM, Jose E. Roman <jroman at dsic.upv.es> wrote:

>
> > El 16/9/2015, a las 21:44, Xujun Zhao <xzhao99 at gmail.com> escribió:
> >
> > The following is another test, which shows the outer iteration is 2, and
> MatMult is performed 24 times. So it means the first outer iteration
> performs 16, and the subsequent iterations performs 8. (24=16+8)
> >
> > What parameters can I tune to accelerate the convergence? Will changing
> ncv work better?
> > The following nconv=1 and unconverged value, what do they mean?
> >
> > 2 EPS nconv=1 first unconverged value (error) 0.0140231 (1.32666590e-02)
> >
> > --->test: EPS solve results:
> >           the eigenvalue: 0.009301
> >           Solution method: krylovschur
> >           Number of iterations of the method: 2
> >           EPS converged reason : 2
>
> 2 outer iterations is very fast, your problem seems very easy. You can try
> changing ncv but I would bet you won’t get the result with much less than
> 24 MatMult’s. Just try.
>
> The monitor you are using shows the number of converged eigenvalues and
> the first non-converged Ritz value.
>
> Jose
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150917/fe878dd4/attachment.html>