Post AW97eEDIs8RpEpGOlE by wakuwaku@mk.absturztau.be
(DIR) More posts by wakuwaku@mk.absturztau.be
(DIR) Post #AW943ajxr6TiWiNZC4 by niconiconi@mk.absturztau.be
2023-05-29T11:07:45.706Z
1 likes, 0 repeats
On accelerating slow physics simulations, this Russian team led by V. Levchenko and A. Zakirov look like a total powerhouse. They've mastered the art of breaking 4D space-time simulation domains into all kinds of crazy geometric shapes for maximum parallelization and cache locality, and they could do it for everything: lattice Boltzmann, Discontinuous Galerkin method, FDTD... Too bad that I can't use any of their studies as they're horrors beyond my comprehension. #hpc
(DIR) Post #AW94ccbutUbaSjJoCO by icedquinn@blob.cat
2023-05-29T11:14:10.944324Z
0 likes, 0 repeats
@niconiconi https://www.mdpi.com/2079-3197/4/3/29 :blobmojiquestion:
(DIR) Post #AW94pRKi8QDXFaNnFo by icedquinn@blob.cat
2023-05-29T11:16:30.921654Z
0 likes, 0 repeats
@niconiconi without having/reading the paper and based on whats in the screenshots, it sorta looks like they just split the world up in to chunks and have some way to rule in neighbor chunks that need to be part of a particular calculation? chunks that have activity that isn't expected to pass over the boundary can be cut and run in parallel?
(DIR) Post #AW95Qtp9AYFsKOTt4K by niconiconi@mk.absturztau.be
2023-05-29T11:23:09.714Z
0 likes, 0 repeats
@icedquinn@blob.cat There's no guidance on exactly how their FDTD domain decomposition procedure works (e.g. how to calculate the coordinates, what does the iterations look like, which is either 3D or 4D and it's very difficult to reason about). All they did is giving the geometric shapes, well, which is sufficient only if you're already an expert on stencil computing, but not me. There's even source code for their early engine on GitHub, but all comments are in Russian, and that was only the 2D+1D time case, not the 3D+1D time case.In fact there are only two teams who've published in the open literature on this technique, another Japanese team led by Fukaya, T. is much more detailed, even with some visualization, but it's still very difficult to reason able with basic questions unanswered, leaving them as an exercise for the reader.Perhaps I can eventually understand it after one year.
(DIR) Post #AW96Pa1zNBqGbdQGNk by niconiconi@mk.absturztau.be
2023-05-29T11:34:14.001Z
0 likes, 0 repeats
@icedquinn@blob.cat Not that simple, this is only a rough sketch in the lower dimension, not the full 3D+1D case. And even then, do note that one of the axis in the graph is not space but TIME, it's not a matter of cutting the array into sections, but instead to do value update in asynchronously in a jigsaw pattern, each cell with its own time state. So at any given time, the entire space is desynchronized, but it will eventually close together at the end with a proper "cleanup" step. The Japanese's team (Takeshi Fukaya)'s paper has more detailed visualizations on a similar partitioning technique.
(DIR) Post #AW96VrT3SPd4QE4s4m by niconiconi@mk.absturztau.be
2023-05-29T11:35:18.820Z
0 likes, 0 repeats
@icedquinn@blob.cat Not that simple, this is only a rough sketch in the lower dimension, not the full 3D+1D case. And even then, do note that one of the axis in the graph is not space but TIME, it's not a matter of cutting the array into sections, but instead to do value update asynchronously in a jigsaw pattern, each cell with its own time state. So at any given time, the entire space is desynchronized, but it will eventually come together at the end with a proper "cleanup" step, difficult to reason about... The Japanese's team (Takeshi Fukaya)'s paper has more detailed visualizations on a similar partitioning technique.
(DIR) Post #AW97eEDIs8RpEpGOlE by wakuwaku@mk.absturztau.be
2023-05-29T11:46:29.669Z
0 likes, 0 repeats
@niconiconi We should build casual browser games that teach you highly abstract concepts. Suddenly everyone would be an expert on machine learning, problems in n-dimensional geometry and maybe even electronics.
(DIR) Post #AW97eEriRsm3GAAfke by niconiconi@mk.absturztau.be
2023-05-29T11:48:03.042Z
0 likes, 0 repeats
@wakuwaku If someone writes a 3D visualization in WebGL on this partitioning scheme and highlights all data dependencies when you select a cell, I'm certain that I can understanding it soon.
(DIR) Post #AW9ABB8TXJZl9mqRjk by ignaloidas@not.acu.lt
2023-05-29T12:16:26.744Z
0 likes, 0 repeats
@niconiconi@mk.absturztau.be making time just another dimension and then doing 4D geometry on it all is something elsepure witchcraft
(DIR) Post #AW9AagT2BQQJZNcgmO by niconiconi@mk.absturztau.be
2023-05-29T12:21:03.678Z
0 likes, 0 repeats
@ignaloidas@not.acu.lt describing it in procedural programming terms, it just means "avoid memory loads at all cost, calculate to as far as you can with the values started with, even if it means the world is broken apart and desynchronized - we can fix that at the end", a pretty straightforward concept. But the iteration pattern and data dependencies are hard to reason about. :woozy_baa:
(DIR) Post #AW9Asi1N92anbLHNse by ignaloidas@not.acu.lt
2023-05-29T12:24:18.932Z
1 likes, 0 repeats
@niconiconi@mk.absturztau.be the fact that it's easier to do 4D geometry than to reason about it in traditional ways tells how hard it is to reason about
(DIR) Post #AW9BDIjoLIJfGfstIe by niconiconi@mk.absturztau.be
2023-05-29T12:28:02.563Z
0 likes, 0 repeats
@ignaloidas@not.acu.lt A related field is polyhedral transformation for compiler optimization.
(DIR) Post #AW9BFzgNe7nh2OZ0dc by ignaloidas@not.acu.lt
2023-05-29T12:28:31.622Z
0 likes, 0 repeats
@niconiconi@mk.absturztau.be sounds cursed