[petsc-users] V-cycle multigrid with matrix shells

[Sanjay Govindjee]

Sylvain,

Is there a reason you are not using Prometheus (--download_prometheus=1) for your MG preconditioner?
It was designed with 3D solid mechanics in mind.

-sanjay


On 5/11/11 11:12 PM, Sylvain Barbot wrote:
> Dear Jed,
>
> During my recent visit to ETH, I talked at length about multi-grid
> with Dave May who warned me about the issues of large
> coefficient-contrasts. Most of my problems of interest for
> tectonophysics and earthquake simulations are cases of relatively
> smooth variations in elastic moduli. So I am not too worried about
> this aspect of the problem. I appreciate your advice about trying
> simpler solutions first. I have tested at length direct solvers of 2-D
> and 3-D problems of elastic deformation and I am quite happy with the
> results. My primary concern now is computation speed, especially for
> 3-D problems, where i have of the order 512^3 degrees of freedom. I
> was planning to test Jacobi and SOR smoothers. Is there another
> smoother you recommend for this kind of problem?
>
> Thanks,
> Sylvain
>
> 2011/5/11 Jed Brown<jed at 59a2.org>:
>> On Wed, May 11, 2011 at 04:20, Sylvain Barbot<sylbar.vainbot at gmail.com>
>> wrote:
>>> I am still trying to design a
>>> multigrid preconditionner for the Navier's equation of elasticity.
>> I have heard, through an external source, that you have large jumps in both
>> Young's modulus and Poisson ratio that are not grid aligned, including
>> perhaps thin structures that span a large part of the domain. Such problems
>> are pretty hard, so I suggest you focus on robustness and do not worry about
>> low-memory implementation at this point. That is, you should assemble the
>> matrices in a usual PETSc format instead of using MatShell to do everything
>> matrix-free. This gives you access to much stronger smoothers.
>> After you find a scheme that is robust enough for your purposes, _then_ you
>> can make it low-memory by replacing some assembled matrices by MatShell. To
>> realize most of the possible memory savings, it should be sufficient to do
>> this on the finest level only.