Re: Sudden random deviation of results

Subject: Re: Sudden random deviation of results

Date: Fri, 8 Dec 2006 15:47:05 +0800

Domainkey-signature: a=rsa-sha1; q=dns; c=nofws; s=beta; d=gmail.com; h=received:message-id:date:from:to:subject:in-reply-to:mime-version:content-type:references; b=n/VWsXAnEC8Xi4AjeKANFRNXKvrOpoaPsZ5h9ghMMTlmhXkCCQjnzOtQyfqXPxGku06O94uFiBDyPhBMbgFag8MfJ6GBZ9F+BS40K8gRMwEoq5+oMLUkVbKPfm/YRnmO7NKBfAnzAHtTiAfhR7pU51kQe4rTDmEOX93Q0850LJ4=

In-reply-to: <Pine.OSX.4.64.0612070832450.212@bsmith-56.mcs.anl.gov>

References: <804ab5d40612070542u57cac6cdudc1fb7272c4f9e59@mail.gmail.com> <Pine.OSX.4.64.0612070832450.212@bsmith-56.mcs.anl.gov>

Reply-to: petsc-users@xxxxxxxxxxx

Sender: owner-petsc-users@xxxxxxxxxxx

Hi,

I have tested my code again, using the checks suggested by Barry. Using KSPGetConvergedReason shows that it has converged. Changing the tol from 1.e-5 to -6 or -7 or even -4 eliminated the explosion of value at that particular time step.

However, that is expected since the values are now different. Moreover, for other linear solver package, it does not happen at this particular time step as well.

Another strange thing is that suppose the explosion of value happen at t=50. If I store my all my relevent values at t=40 into files and restart my simulation from t=40(instead of t=1) by reading the stored values from the files, the explosion will now not happen at t=50. Hence it seems very random too....

On 12/7/06, Barry Smith <bsmith@xxxxxxxxxxx> wrote:

Ben,

  First make sure that the residual has actually gotten
to the tolerance you want. PETSc KSP does NOT stop if the
linear system has not converged; you should call KSPGetConvergedReason()
after each solve to make sure it has converged (as a quick check
you can run with -ksp_converged_reason.)

Next try decreasing the KSP tolerance factor a great deal. Does
1) the solution you get for your physics problem look the same for small
  time (after a few time-steps they will start to drift).
2) the residual blow up at the same time, or later?

  Barry

On Thu, 7 Dec 2006, Ben Tay wrote:

> Hi,
>
> I have been using a few different Krylov linear solver package such as
> nspcg,sparsekit and now petsc to solve the linear eqns for my NS solver
> momentum and poisson eqn. For moving bodies simulation, while solving the
> poisson eqn, sometimes the result suddenly changes drastically, although the
> iterations seems to have converged e.g. at the previous time step,
> everything is fine, but at the next time step, the velocity suddenly
> explodes from 2-3 to 20-30+.
>
> The strange thing is that this deviation happens at different time step for
> different linear solver package ie for nspcg, it happens at time=20 but for
> petsc, it happens at time=40. It seems "random" in this sense. Changing the
> type of solver such as gmres or bcgs does not change the time of deviation.
>
> Does anyone know why this is happening? Is it a characteristic of krylov
> solver to happen once in a while? It does not happen in the stationary body
> simulation though.
>