[Carpet] Tapered grids
Erik Schnetter
schnetter at cct.lsu.edu
Fri May 5 16:33:27 CEST 2006
On May 5, 2006, at 08:17:36, Ian Hawke wrote:
> Dear all,
>
> For those that like playing with totally untested patches, I attach
> the
> following (darcs DEV version). As the description says, it _should_ do
> the tapered grid approach, but you'll need to specify the number of
> buffer zones by hand. I haven't tested much (except to check that the
> default behaviour of Carpet is unchanged).
If you want to use tapered grids to avoid the storage for past time
levels, consider the following:
Assume that you have 30 evolved grid functions, 8 levels, and each
level has 50^3 grid points. Assume also that you use RK4 as time
integrator and a 4th order centred method, requiring 2 ghost zones.
This requires 4 x 2 = 8(4 integrator substeps, 2 stencil size) buffer
+ ghost zones for Carpet's current approach, and 2 x 8 = 16 buffer +
ghost zones for a fully tapered approach. Currently, you need 3 time
levels plus the RHS, with a fully tapered approach you need only 2
time levels plus the RHS.
The total amount of storage for both cases is then:
current Carpet:
30 x 4 x (50 + 2x8)^3 x 8 = 275,996,160 Bytes
tapered:
30 x 3 x (50 + 2x16)^3 x 8 = 396,984,960 Bytes
That is, you will need more memory, not less. The ratio changes in
favour of a fully tapered approach if you use larger grids or lower
order methods. The break even occurs e.g. approximately for 100^3
grids (which are very large) and an RK3 integrator instead of RK4.
Note that many people use a stencil size of 3 instead of 2 (which I
assumed), which makes things worse again.
-erik
--
Erik Schnetter <schnetter at cct.lsu.edu>
My email is as private as my paper mail. I therefore support encrypting
and signing email messages. Get my PGP key from www.keyserver.net.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: PGP.sig
Type: application/pgp-signature
Size: 186 bytes
Desc: This is a digitally signed message part
Url : /archives/developers/attachments/20060505/99e11024/attachment.pgp
More information about the developers
mailing list