4 ms·
So based on my quickie checks... is this a task-based dynamic scheduling (as opposed to static thread-based scheduling) that extends out to CUDA / OpenMP / othe
by dragontamer 7y ago
So based on my quickie checks... is this a task-based dynamic scheduling (as opposed to static thread-based scheduling) that extends out to CUDA / OpenMP / other backends?
Which sounds pretty good to me. I'm writing a very basic version of future + task based scheduling for Chess AI, so I think this methodology is very powerful. I haven't dug very deeply into your work, but it seems like you guys have a good idea of what you're doing.
Seems like a cool project. Unfortunately, my home computer is AMD Vega64, so I won't be able to use Legion. But its still got a methodology that I personally think, is really cool.
---------
The silly example that "sold" me on task-based scheduling was Intel's "Stupid Fibonacci" example. I think you guys should put a similar example on your website: https://software.intel.com/en-us/node/506102 https://software.intel.com/en-us/node/506102
Stupid parallel Fibonacci is basically fizzbuzz, every programmer should understand it. And while its not very useful, you should be able to represent it in Legion in less than 10 lines of code.
EDIT: I see that your code actually has Fibonacci. Good :-) https://legion.stanford.edu/tutorial/tasks_and_futures.html https://legion.stanford.edu/tutorial/tasks_and_futures.html
Well, I have nothing to contribute. Good stuff.
- eslaught 7y agoIt's shorter in Regent :-) https://github.com/StanfordLegion/jupyter-regent/blob/master/notebooks/Introducing%20Regent.ipynb https://github.com/StanfordLegion/jupyter-regent/blob/master... But anyway, the reason we don't really push fibonacci is because it (a) doesn't really expose the unique features of Legion/Regent and (b) if you write it naively, probably performs even worse than naive fibonnacci implementations on other parallel frameworks. The real secret sauce in Legion is being able to load a mesh or whatever, partition it a couple different ways (e.g. for ghost cells vs the main computation) and then launch tasks without needing to think about what needs to be updated where to get the data to the right place or copying data down to the GPU and back. We will add AMD GPU support eventually, since Frontier is going to be a GPU machine, but probably not on a time scale that makes any difference to you.