Talk:Checkpoint synchronization: Difference between revisions
Content added Content deleted
(I like FFT, but not specifically for this.) |
(What it does, what it might be used for) |
||
| Line 4: | Line 4: | ||
::: FFT would be very interesting, because it's a very popular signal processing transform, and there are huge number of ways to improve performance via platform instruction sets, acceleration libraries and parallelization. It's very likely it will lead to multiple examples per language, but as long as there's at least one naive example per language (for logic clarity), I think that's probably a good thing. |
::: FFT would be very interesting, because it's a very popular signal processing transform, and there are huge number of ways to improve performance via platform instruction sets, acceleration libraries and parallelization. It's very likely it will lead to multiple examples per language, but as long as there's at least one naive example per language (for logic clarity), I think that's probably a good thing. |
||
::: That said, I'd be careful about appropriating FFT specifically to demonstrate "[[Checkpoing synchronization]] with a communication phase"...for something like that, a more appropriate task might be something like map/reduce with an otherwise meaningless function call on each data block. (I wanted to say "generic function call", but that might force a distinction between languages that have (or can fake) first-class functions, and languages that don't/can't.) I like the FFT idea, but I don't care to have distinct tasks for a FFT map/reduce-like task, and FFT in general. Ditto for matrix block multiplication, for the same reasons; it's not helpful for site organization to demonstrate the same thing (such as a transform or algorithm) in multiple places. --[[User:Short Circuit|Michael Mol]] 10:46, 11 August 2010 (UTC) |
::: That said, I'd be careful about appropriating FFT specifically to demonstrate "[[Checkpoing synchronization]] with a communication phase"...for something like that, a more appropriate task might be something like map/reduce with an otherwise meaningless function call on each data block. (I wanted to say "generic function call", but that might force a distinction between languages that have (or can fake) first-class functions, and languages that don't/can't.) I like the FFT idea, but I don't care to have distinct tasks for a FFT map/reduce-like task, and FFT in general. Ditto for matrix block multiplication, for the same reasons; it's not helpful for site organization to demonstrate the same thing (such as a transform or algorithm) in multiple places. --[[User:Short Circuit|Michael Mol]] 10:46, 11 August 2010 (UTC) |
||
: The principle is that there is a checkpoint group (which a thread may ask to join and leave) and on synchronization, nobody completes the checkpoint synch operation until all members of the group complete it. It's use might be if you've got a bunch of threads doing some kind of physical simulation, with a checkpoint barrier at the end of each simulation time step. –[[User:Dkf|Donal Fellows]] 13:53, 11 August 2010 (UTC) |
|||