[Nek5000-users] Tolerances

[nek5000-users at lists.mcs.anl.gov]

tolrel and tolabs are ineffective when divergence/helmholtz
are nonzero (i.e., they over-ride).

tolrel determines conv. criteria based on eigenvalue estimates
and magnitude of data (hence "rel" for relative).  When tolrel
is in effect, tolabs is ineffective.

tolabs is like tolrel except it works when nek can't determine
the magnitude (i.e., somehow all "observable" data is zero,
where observable is to be interpreted as the data that tolrel
has been designed to trigger off of).

Typically I prefer div/hmh for problems that are nondimensional
by convective time scales (i.e., where visc == 1/Re).  In those
cases, 1.e-6 for div and 1.e-8 for hmh are pretty good for most
apps.    Exceptions are stability problems or convergence studies,
where tighter tolerances are desirable.

Most likely you can't go too far wrong w/ tolrel=1.e-9 and
hmh/div=0,  though I find tolrel is a bit conservative.  (Incidentally,
engineering tolerances for tolrel would typically be 1.e-2, and
would be very satisfactory -- in fact a bit more conservative than
the 1.e-6/1.e-8 figures of the preceding paragraph.)

The other option would be to set div=1.e-9, hmh=1.e-10 or 11.
Assuming you're working in convective timescales, these should
be adequate for your stability problem.

15 iterations doesn't seem to be a lot --- One question, do you
have dt fixed (param 12 < 0), and p93, 94, 95 > 0 ?   The latter
control the pressure / velocity projection, which can greatly
accelerate performance, esp. for tight tolerances.  It is most
effective when dt = constant.

hth,

Paul





On Wed, 28 Apr 2010, nek5000-users at lists.mcs.anl.gov wrote:

> Dear Paul and Stefan,
>
> We are currently studying a 3D flat-plate boundary layer which is highly 
> unstable to crossflow modes. When computing the base flow we realize that 
> numerical noise already triggers unsteady crossflow waves with amplitudes of 
> the order of 10^{-6}. We plan to study receptivity to external perturbations 
> with amplitudes of 10^{-5} (to avoid early nonlinearity) so it would be 
> desirable to suppress the numerically triggered disturbance waves.
>
> We use the PN-PN-2 formulation, dealiasing and filtering (0.05). We also 
> noticed that about 15 pressure iterations per timestep (gmres) were needed 
> which seems to be high.
>
> We further use the following tolerances:
>      1.0E-08         DIVERGENCE
>      1.0E-08         HELMHOLTZ
>
>      1.0E-06         TOLREL
>      1.0E-10         TOLABS
>
> We wonder whether using stricter values may lead to smoother results. What is 
> actually the meaning of TOLREL and TOLABS?
>
> Thank you very much for your advice and some clarification.
>
> Greetings from KTH Stockholm,
>
> Lars-Uve and David
>
>
>
> _______________________________________________
> Nek5000-users mailing list
> Nek5000-users at lists.mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/nek5000-users
>