[Sidefx-houdini-list] Distributed Fluid Sim question

John Coldrick jc at axyzfx.com
Tue Aug 11 10:07:39 EDT 2009


On Tuesday 11 August 2009 09:04, Beatriz Lorenzo wrote:
> I think it depends on the sim.

	Absolutely.  There is no single rule.  Testing is in order.  ;)

> As you need an overlapping region for 
> distributing the voxel regions, if the sim is very complex and fluid is
> moving in many directions (=needs a lot of feedback from the adjacent
> voxels) it might end up in lots of artifacts if you distribute in too many
> jobs.

	Depending on just *how* complicated it is, busy-ness can often hide 
artifacts, too.  I've found it's the relatively still ones that give me 
trouble more often than not, although of course there are dozens of variables 
here.  Many sims simply *cannot* behave properly when divied up, it's 
inherent in what you're trying to do.  The division walls can become factors 
which actually profoundly affect the sim.  As far as efficiency, though, the 
more machines you run on, the more crosstalk you get/need, although the 
actual volume of data isn't really all that bad, with many machines thrown at 
it, you will be getting a bit of a hit with that.  Of course, whether this 
balances out the extra cycles...again, testing is in order.  :)

	When it comes right down to it - the most *efficient* scenario is a single 
machine with stupid amounts of ram and a ridiculous number of threads thrown 
at it.  Alas, this isn't exactly the most common render farm scenario...it's 
up to 16 threads now with dual 8-core i7's, but complex sims can require 
truly colossal amounts of ram.

	Cheers,

	J.C.

-- 
John Coldrick                  www.axyzfx.com        Axyz Animation
416-504-0425                                         477 Richmond St W
                                                     Toronto, ON Canada
jc at axyzfx.com                                        M5V 3E7
-----------------------------------------------------------------------
Living hell is the best revenge.
  - Adrienne E. Gusoff




More information about the Sidefx-houdini-list mailing list