Multi-Grid for Rectangular Domain


RGW6
03-05-2003, 07:36 PM
I need to use mglin in my dissertation work on a rectangular grid. I'm just about ready to start converting mglin.for for use on a rectangular grid.....are there any particular areas that I should be especially careful about? Has anyone attempted this before?

Thanks,

Gary

RGW6
03-07-2003, 07:36 AM
I just about have all the changes made so that mglin can work on a rectangular grid. I'm setting it up for a 1:10 grid. My coarsest grid will be 3-by-21 (x-to-y) and I'm keeping dx=dy.

I've been debating (with myself) about how to get the "exact" solution for the coarse grid.... a direct method or just use an iterative method like SOR. Right now I'm planning to just use SOR.....any thoughts here??

Also, does anyone know where the equation for MEMLEN originates?

hernlund
03-18-2003, 02:12 AM
When doing multi-grid on any domain, one must choose the mesh as multiples of 2 in order to get the best perfomance and take the coarse grid corrections to the finest grid. So you should always use 2,4,8,16,32,64,128,256, etc. in any direction. Multi-grid convergence can be quite fussy even when you change a simple parameter of your model, so one has to be careful and really analyze the convergence on all grids. The chapter on this topic in NR is a good one, and gives a lot of philosophical guidance, but if your problem is not so simple then you may have to consult the literature. The code given in NR is just a very basic one and will need a lot of modification for most problems. Some multi-grid packages are available from other sources (eg www.mgnet.org), although these are not optimized for specific problems since they are quite general, and so there is a trade-off...if your problem is on a very small grid (i.e. on the order of 10s in each direction) then multi-grid is not really necessary. It is only when you get into hundreds of thousands or millions of grid points that this is a useful technique.