Re: [Algorithms] General purpose task parallel threading approach

[Sebastian S. <seb...@gm...>]

On Sat, Apr 11, 2009 at 9:59 PM, Jarkko Lempiainen <al...@gm...>wrote:

>  Task priorities are updated once per frame, but there is no list of tasks
> to be sorted before execution of tasks begins. That’s purely your own wrong
> conclusion drawn from… wherever.
>
For example the presentation you posted that (as far as I can tell without
seeing the talk) talked about static ordering of tasks ahead of time. And
you also talked about global prioritization, and how you weren't waiting on
other tasks within a task etc.

> So, it’s online scheduler in the sense that tasks are spawn “randomly”
> within tasks, but their priorities are determined in the end of the frame
> for next execution based on task dependencies and execution times. That
> clears things up?
>
Okay, that makes sense.
You're not actually resolving dependencies by sorting the tasks up front,
dependencies are still enforced dynamically, you just sort things that do
not have dependencies based on historical behaviour? I.e. you spawn a task
and *if* that task existed last frame you prioritize it based on that?
Have you measured if the added communication overhead of doing this actually
buys you anything on real code? Seems like you'd have to make sure that the
overhead of moving things around based on priority must be counteracted by
the benefits, and since the constant factor for work-stealing is very close
to 1 in empirical studies there doesn't seem to be that much benefit to be
gained from shuffling tasks around...

If not, please ask rather than draw further wrong conclusions.
>
I usually try hard to stay on topic, but seriously, I asked you multiple
concrete questions about how your system would handle specific scenarios,
answering one of those would probably have been more illustrative than
throwing ad-hominems around...

-- 
Sebastian Sylvan
+44(0)7857-300802
UIN: 44640862