Feedback

Tag cloud

What's your question?

By: [ Editor ] Asked

Renderfarming on computers, processors or cores?

There are various renderfarm task / processor configuration strategies, probably the most common being the following three:

  • One task for a machine, regardless of processor and core count. The executed software is supposed to do all the threading stuff for itself.
  • One task for a processor, so multiprocessor machines run multiple tasks at the same time on all cores of the proc.
  • One task for a core, that results in lots of tasks running simultaneously but probably avoiding thread locks.

All of these configurations have some serious pros and cons (mostly related to licensing, network IO, memory usage and threading capability of the softwares) and I'd like to hear your experiences, ideas about them.

Add comment viewed 166 times Latest activity 7 months ago

NN comments
or Cancel

2 answers

  • 3

paul nendick [ Editor ]

The thought process for me and mine goes thusly:

We don't typically make sweeping changes to our pool of render resources very often so the hardware characteristics are effectively constant.

The choice to purchase a given quantity of licenses of a given piece of software is made with budget considerations more than for performance ones. So quantity of licenses is effectively constant.

Software vendors have always and will always make the licensing terms suit their needs and not ours so per-core/per-host/per-site/node-locking options are out of our control as well.

The only variables left are how one sub-divides the pool of hardware resources and then how a software maps to a unit of those resources.

So I don't make a farm-wide strategy, instead I make a per-software tool strategy. Quantify the performance of each piece of software on each known hardware profile using each plausible software configurations. Let that inform the making of an optimised launch strategy for each tool to then be asserted in the render job launching process.

NN comments
szabolcs
-

This looks like a good general overview of the thought process. It would be great to hear some actual experiences with various configurations though. Profiling such things is not that straightforward (taking all software and task type combinations on the same computer into account) so any first hand experience would be interesting.

or Cancel
  • 0

vfxoverflow_168 [ Editor ]

I talked with a RenderMan coder and he told me that usually one frame is computed per core (so, no multithreading etc..). It is convenient when you render an animation : the cost of rendering one frame or rendering 5000 is the same (12 hours for Avatar for example), you just use 5000 cores in the latter case.

NN comments
or Cancel

Question's Contributors

Related Questions