On Wed, Aug 25, 2010 at 6:22 AM, Michael Koppitz koppitz@aei.mpg.de wrote:
Hi all,
I made the observation that Carpet's output is quite large compared to the grid structure:
on 100^3 gridpoints per refinement level dumping admbase::metric (6 gridfunctions) in 3d hdf5
I would expect for 1 ref-level 10^6*6*8b = 48Mb. I got 80Mb. I assume that is because Carpet also includes x, y, z, and t in the dump, which brings it to 10 GF's and 80Mb expected.
But for 2 levels I got 250Mb. For 3 levels I got 450Mb and for 4 levels I got 620Mb.
This seems much even if I take ghost- and buffer-zones into account.
Am I missing something, or is Carpet doing something unexpected?
If you have a large number of ghost zones, then these require a surprising amount of memory. This could explain the difference between 48MB and 80 MB.
On the finer levels there will also be buffer zones. These are also expensive. This may explain the numbers.
You can look at the hdf5 files and check the size of the datasets in there. There will also be a bit of administrative information in the hdf5 files, but this should not be that much.
-erik