I recently did some volume renders of a 50 Mpc box unigrid simulation with 1024^3 grid cells on kraken. I used exactly 64 cores and did not have to use less than the full number of cores available per node. I was making 1024^2 images that took roughly between 5-10 seconds to render. I tried some 2048 that took around 30-40 seconds. I was rendering baryon overdensity with a transfer function that had 2000 narrow gaussians. The number was high because I am combining this with a movie in which I render only one of those guassians at a time and build the box up from low overdensity up to high. I didn't go to lower number of processors, so I'm not exactly sure at what point this would have run out of ram. I consider this an overwhelming success. I've attached some sample images, one with the full transfer function and a sample frame from the movie where I do them one at a time while spinning. Very very nice job!
Great work! I'm really happy to see this make it into the primary trunk.
I'd like to encourage people to try this out, particularly on large
datasets, and write to the list or Sam if you run into problems. This
is a big increase in functionality, and everyone wants to make sure it
works out alright.
I've been using the volume rendering capabilities of yt quite
extensively, in kind of an unconventional way, to calculate off-axis
average values, and I'm very excited about the performance
improvements that this new subsystem will bring.
On Tue, Nov 9, 2010 at 5:12 PM, Sam Skillman <firstname.lastname@example.org> wrote:
> Hi all,
> I just wanted to announce that the new kd-Tree rendering framework is now in
> the 'yt' branch of the repository. There are a couple things I wanted to
> point to if you are interested.
> The changeset itself:
> A post on blog.enzotools.org highlighting some recent successes:
> A simple script, where you should just have to change the parameter file
> A more advanced script that exposes a few new options:
> Both of these scripts should be able to be run in parallel (as long as N is
> a power of 2 for now) transparently as:
> mpirun -np N python script.py --parallel
> Parallel performance will depend on the structure of your data, but the docs
> for the Camera object have some suggestions.
> If you find any problems or have any thoughts, let me know!
> yt-users mailing list
yt-users mailing list