By: Ireland (boh.delete@this.outlook.ie), January 26, 2017 5:09 pm
Room: Moderated Discussions
RichardC (tich.delete@this.pobox.com) on January 26, 2017 3:24 pm wrote:
>
> And that's really where this discussion started out: the claim was that all
> "supercomputer" apps required a reasonably high network bandwidth/compute ratio.
> And I suggested high-quality 3D rendering as a significant app which can use
> a huge amount of compute, is easily distributable across a cluster of shared-nothing
> machines (even a cluster of unreliable machines), and requires only very low
> network bandwidth to/from each machine (though very possibly high bandwidth for
> some of the ways you might want to use the rendered frames).
>
> I rest my case.
>
Where you'll really figure this out, is by looking at how Pixar charges for those licenses to use renderman based on how many CPU sockets that the rendering production company intends to use. I mean, Pixar knew about the concept a long, long time ago - about ganging together separate boxes in order to chew through the task of rendering. However, even their own estimates for the 'high' number of boxes and CPUs that would be involved - had to be upgraded only two years ago - so obviously, one can take from that, that this renderman thing scales very well across even larger numbers of CPU's.
It was interesting too, that the Hyperion system that Disney were using in 2014-ish to do their 'Hero 6' feature length movie - they didn't actually refer to it being a 'rendering farm', or a cluster of machines - as we would have done long ago. Instead, they actually refer to their hardware now, just as a 'super computer'. I think that maybe this is where Aaron is coming from, in his emphasis in this conversation - upon the 'super' as opposed to the 'cluster. So obviously, there is some change of mindset happening down at the level of these movie production companies, and how they view their process, compared to the past. I don't know. It was interesting too I thought, how the Hyperion system with the 55,000 cores were distributed across four geographical locations, and they only used management software to make it appear like one large super computer - when actually it was big banks of processors assembled together - in four different locations.
Maybe that enables them to work on separate projects on the four different clusters, during 'off peak' times, when developing the concepts - and then when it gets closer to production quality stages of a project - they're able to group these four systems together, and make them all burn like one large furnace. That would seem to support my idea, that one needs a way to break these systems down into smaller systems - to attack problems at a smaller scale - and yet have a capability to harness everything into one 'warp speed' kind of mode too (i.e. Scottie, down in the boiler room, I need you to give me all of the 'steam' that you can generate).
>
> And that's really where this discussion started out: the claim was that all
> "supercomputer" apps required a reasonably high network bandwidth/compute ratio.
> And I suggested high-quality 3D rendering as a significant app which can use
> a huge amount of compute, is easily distributable across a cluster of shared-nothing
> machines (even a cluster of unreliable machines), and requires only very low
> network bandwidth to/from each machine (though very possibly high bandwidth for
> some of the ways you might want to use the rendered frames).
>
> I rest my case.
>
Where you'll really figure this out, is by looking at how Pixar charges for those licenses to use renderman based on how many CPU sockets that the rendering production company intends to use. I mean, Pixar knew about the concept a long, long time ago - about ganging together separate boxes in order to chew through the task of rendering. However, even their own estimates for the 'high' number of boxes and CPUs that would be involved - had to be upgraded only two years ago - so obviously, one can take from that, that this renderman thing scales very well across even larger numbers of CPU's.
It was interesting too, that the Hyperion system that Disney were using in 2014-ish to do their 'Hero 6' feature length movie - they didn't actually refer to it being a 'rendering farm', or a cluster of machines - as we would have done long ago. Instead, they actually refer to their hardware now, just as a 'super computer'. I think that maybe this is where Aaron is coming from, in his emphasis in this conversation - upon the 'super' as opposed to the 'cluster. So obviously, there is some change of mindset happening down at the level of these movie production companies, and how they view their process, compared to the past. I don't know. It was interesting too I thought, how the Hyperion system with the 55,000 cores were distributed across four geographical locations, and they only used management software to make it appear like one large super computer - when actually it was big banks of processors assembled together - in four different locations.
Maybe that enables them to work on separate projects on the four different clusters, during 'off peak' times, when developing the concepts - and then when it gets closer to production quality stages of a project - they're able to group these four systems together, and make them all burn like one large furnace. That would seem to support my idea, that one needs a way to break these systems down into smaller systems - to attack problems at a smaller scale - and yet have a capability to harness everything into one 'warp speed' kind of mode too (i.e. Scottie, down in the boiler room, I need you to give me all of the 'steam' that you can generate).