Hi yt-dev, I am developing a frontend for Einstein Toolkit for yt and in the process I generated some crude, preliminary benchmarks which I thought I would share in case anybody is interested. I performed three tests: 1. I just load the dataset and calculate (say) the maximum of some quantity on each grid. 2. I load in the dataset and calculate the maximum of the magnitude of a gradient on each grid. This requires the generation of ghost zones at grid boundaries. 3. I load in the dataset and perform a volume rendering. I make a "movie" with 4 frames where I rotate around the volume. Salient details: * I performed these tests with the attached scripts. I turned openmp and MPI off (set OMP_NUM_THREADS=1). * The domain sizes were: 64^3, 128^3, 256^3, and 512^3. * All four datasets have only one refinement level. However the 512^3 dataset has multiple grids on that refinement level. * The size of a dataset ranges from about 10MB for the 64^3 set to 10GB for the 512^3 * All data was performed on a single modern workstation. Details: 2.6GHz clock with AVX2 instructions and a 60MB L3 cache. Results: * Reading in the data is (for the datasets I tested) extremely fast. The 512^3 dataset takes only about 20 seconds to read the whole thing in. * Generating the first frame in a volume render is extremely slow. On the order of 4 minutes for the 512^3 data set. After the first frame is produced, new frames are fast, even with openmp off. With 1 openmp thread, it takes on the order of 10s of seconds for a new frame. * Generating ghost zones is very fast for datasets with only one grid. It is incredibly slow for datasets with multiple grids, dominating the run time. * I attach a plot comparing the three tests. Naively, it seems to me that there must be several sources of overhead when I perform volume rendering that are significantly more costly than simply reading in the data (which seems to be fast). Clearly one of these is the generation of the ghost zones, and another is the actual ray tracing (although the ray tracing itself seems to be quite fast.). However I'm not sure that these two operations alone explain the cost of the volume rendering. I'd love to know other people's experience with benchmarking. Do these costs seem normal, up to an order of magnitude or so? Would you have any insight into what contributes to the cost of generating the first frame when volume rendering? Thanks very much! Best, Jonah Miller
Hi Jonah, Thanks for supplying these. I've reviewed them but am still processing. On Tue, May 3, 2016 at 10:48 PM, Jonah Miller < jonah.maxwell.miller@gmail.com> wrote:
Hi yt-dev,
I am developing a frontend for Einstein Toolkit for yt and in the process I generated some crude, preliminary benchmarks which I thought I would share in case anybody is interested.
I performed three tests:
1. I just load the dataset and calculate (say) the maximum of some quantity on each grid. 2. I load in the dataset and calculate the maximum of the magnitude of a gradient on each grid. This requires the generation of ghost zones at grid boundaries. 3. I load in the dataset and perform a volume rendering. I make a "movie" with 4 frames where I rotate around the volume.
Salient details:
- I performed these tests with the attached scripts. I turned openmp and MPI off (set OMP_NUM_THREADS=1). - The domain sizes were: 64^3, 128^3, 256^3, and 512^3. - All four datasets have only one refinement level. However the 512^3 dataset has multiple grids on that refinement level. - The size of a dataset ranges from about 10MB for the 64^3 set to 10GB for the 512^3 - All data was performed on a single modern workstation. Details: 2.6GHz clock with AVX2 instructions and a 60MB L3 cache.
Results:
- Reading in the data is (for the datasets I tested) extremely fast. The 512^3 dataset takes only about 20 seconds to read the whole thing in.
That's good news!
- Generating the first frame in a volume render is extremely slow. On the order of 4 minutes for the 512^3 data set. After the first frame is produced, new frames are fast, even with openmp off. With 1 openmp thread, it takes on the order of 10s of seconds for a new frame.
That sounds about right; my guess is that this is the kdtree building, which will get the vertex centered data.
- - Generating ghost zones is very fast for datasets with only one grid. It is incredibly slow for datasets with multiple grids, dominating the run time.
Sounds about right.
- I attach a plot comparing the three tests.
Naively, it seems to me that there must be several sources of overhead when I perform volume rendering that are significantly more costly than simply reading in the data (which seems to be fast). Clearly one of these is the generation of the ghost zones, and another is the actual ray tracing (although the ray tracing itself seems to be quite fast.). However I'm not sure that these two operations alone explain the cost of the volume rendering.
I think that the cost is likely dominated by the vertex-centering. It may be possible to overload get_vertex_centered_data for your subclass of AMRPatchGrid to make this faster based on your data format. I am not sure that it is a huge cost to do the ray casting itself; there are likely python operations like writing the image, etc, that are a non-negligible fraction of that time.
I'd love to know other people's experience with benchmarking. Do these costs seem normal, up to an order of magnitude or so? Would you have any insight into what contributes to the cost of generating the first frame when volume rendering?
Thanks very much!
Best, Jonah Miller
_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
Hi Matt, Thanks very much for your comments! On 16-05-06 11:58 AM, Matthew Turk wrote:
I think that the cost is likely dominated by the vertex-centering. It may be possible to overload get_vertex_centered_data for your subclass of AMRPatchGrid to make this faster based on your data format. I am not sure that it is a huge cost to do the ray casting itself; there are likely python operations like writing the image, etc, that are a non-negligible fraction of that time.
I see. Do you have any insight into how we might prepare the data to facilitate this? Or how we might prepare the data to minimize the cost here? For example, is yt more efficient with many small grids or with a few large grids? Since we perform a pre-processing step before feeding our data to yt, it may be possible to generate grids that are efficient for yt. Best, Jonah
I'd love to know other people's experience with benchmarking. Do these costs seem normal, up to an order of magnitude or so? Would you have any insight into what contributes to the cost of generating the first frame when volume rendering?
Thanks very much!
Best, Jonah Miller
_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org mailto:yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
participants (2)
-
Jonah Miller
-
Matthew Turk