Re: [Algorithms] General purpose task parallel threading approach

[Jarkko L. <al...@gm...>]

> It's actually not vital for throughput, and I couldn't see anything in the
presentation you linked saying it is.

 

That's what "Minimizing the Critical Path" is about. Ignoring the issue
doesn't mean it doesn't exist.

 

 

Cheers, Jarkko

 

 

  _____  

From: Sebastian Sylvan [mailto:seb...@gm...] 
Sent: Friday, April 10, 2009 1:13 AM
To: Game Development Algorithms
Subject: Re: [Algorithms] General purpose task parallel threading approach

 

 

On Thu, Apr 9, 2009 at 10:27 PM, Jarkko Lempiainen <al...@gm...>
wrote:

Yes, appealing to the "big boys" is pretty weak argument since they have
provided sub-optimal solutions before as well. 

In this case almost all of them have, at one point or another, tried the
alternative and decided it was worth the effort to implement the much
trickier solution. In fact the the presentation you linked was given by the
very same organization (xna) who now recommend work stealing - so clearly
they've given the global work queue an honest shot don't you think?

 

Anyway, I was hoping to get some strong arguments either way, but I guess
it's not very clear yet what solution is the best for multi-threading your
engine. I read through the Gamefest slides and it was a bit surprising they
didn't deal with prioritization at all, which is vital for throughput.

It's actually not vital for throughput, and I couldn't see anything in the
presentation you linked saying it is.

Tasks are naturally ordered according to their dependencies due to the way
they are spawned (you spawn tasks, then you sync - only then do you spawn
any dependent tasks). With the cilk based model any task that is enqueued is
by definition ready to run. There's in no need for priorities to optimize
throughput. 

 

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