[CIG-MC] Re: CitcomCU/CitcomS convergence parameters

[Eh Tan]

Hi Magali,

I think your question might be interesting to many people, so I will 
reply to the mailing list.

These four parameters (down_heavy, up_heavy, vlowstep, vhighstep) 
control the number of smoothing cycles for each multi-grid level. Each 
smoothing cycle is a Gauss-Seidel iteration over the residual and 
involves two global communication.  Apply more smoothing cycles per 
multi-grid level will make the residual decreases faster, but each level 
will also take longer time to finish. Clearly, there is a trade-off that 
depends on your machine configuration: the type of network connection, 
the speed of the CPU, router, etc...

The rule of thumb is that down_heavy, up_heavy and vhighstep should be 
small integers (2 or 3 is usually good enough). But you might need to do 
a benchmark to determine the best vlowstep for your machine AND your 
problem size. For fast network or small problem (small # of processors), 
use higher value of vlowstep. For slow network or big problem (large # 
of processors), use lower value of vlowstep. Shijie told me 20 is 
usually big enough for vlowstep. I also have the feeling that the 
current default value of vlowstep (=1000) in CitcomS is probably way too 
big.



Eh


Magali Billen wrote:

> Hello Eh,
> I've started using CitcomCU, so I can use the multigrid solver 
> (citcomT only has conj-grad) and
> the tracers.  Anyway, this is the first time I'm using the multigrid 
> solver and I want to make sure
> I'm using the input parameters correctly.  The only questions I have 
> are related to whether there
> are 'standard' values for the parameters:
>
> down_heavy
> up_heavy
> vlowstep
> vhighstep
>
> These all seem to effect the smoothing effect with the multigrid 
> solver at different levels (by
> repeating the solutions more than once at each level).
>
> It seems like down_heavy, up_heavy and vhighstep should all be a 
> smallish number like 3.
> Is there a rule of thumb for what vlowstep should be  (200, 1000) or 
> do I just need to run a bunch
> of benchmark tests ?? 
>
> Also, is this the type of question that should be posted to the 
> mailing list?
> Thanks,
> Magali
>
> ---------------------------------------------
> Assistant Professor
> U.C. Davis, Geology
> (530) 754-5696
> billen at geology.ucdavis.edu <mailto:billen at geology.ucdavis.edu>
> ----------------------------------------------
> "Life is not measured by the breaths you take, but by the moments that 
> take your breath away."
>
>