[Beowulf] Digital Image Processing via HPC/Cluster/Beowulf - Basics

Vincent Diepeveen diep at xs4all.nl
Mon Nov 5 17:16:30 PST 2012


Jim as someone who produced games, this is not how it works for most  
movies/animations/commercials where graphics work is needed.

Note that most movies get editted as well in the same way (each  
nation usually has different requirements).

First of all - the budget is rather tiny. So that means using  
hardware you would probably yourself not consider using at NASA.

Secondly the calculation power required to produce movies is rather  
limited. Even a graphics card from years ago has more
capabilities than any artist single handed can design for. The real  
limitation is the amount of graphics a single person can produce.

Producing a head requires the work of an artist specialized in that,  
for a full month. Rendering that is a manner of minutes at
a single core. We speak of a CPU.

A whole animation in full blown HD gets rendered just as quick.

You can render the next episodes of the upcoming starwars movies deep  
at a single socket i7 pretty quick, which includes
testrun after testrun.

Now hollywood and marketing department wouldn't be doing what they  
get paid for if they don't blow that up into gigantic
proportions adding up every workstation any designer once ran onto,  
then squaring that computation power and claiming that
as something they 'used' to produce the next movie.

People are easier to draw to watch a movie saying that you used all  
computation power produced during the past 100 years
to produce this movie, than if you do a 1 line statement: "We kept it  
low budget and rendered the final movie at a single socket i7
involing only our file server as well."

Even then - when we rendered a movie out of a game - consider this.

Our 3d-engine at what by now is a very old junk GPU, he it wasn't  
even the fastest GPU back then - the 3d engine
rendered that actually better and REALTIME, so that took a few  
SECONDS, than the official rendering software which took 30 minutes
of a single core - and some of the colors of the details the  
rendering software produced weren't 100% the same as the designers  
had intended it - our own 3d engine (in OpenGL) was doing it better.

It's always possible to slowdown that rendering you know.

This is not rocket science software. If some professional programmers  
would speedup that rendering engine - it for sure will get
a factor 100 faster handsdown.

I bet there isn't enough budget to even port that rendering engine of  
the design software to be using gpgpu support...




On Nov 5, 2012, at 11:49 PM, Lux, Jim (337C) wrote:

> What I find interesting (and which is characteristic of this  
> application space) is the sort of bimodal requirement:
>
> 1)      High performance workstation with bespoke software that the  
> digital artist uses
>
> 2)      A render farm to grind out the final product.  (quite the  
> EP task, in general)
>
>
>
> The workflow is similar to the traditional film workflow, where  
> each person gets a piece to work on, and then it’s handed off to  
> someone else to composite with the other pieces and build up the  
> whole film.  The artist would work in wireframe or with rendered  
> key frames, do the changes, then send it off to be rendered.  The  
> next work day, the fully rendered product is complete and viewable.
>
>
>
> The other interesting thing is that this problem space has HUGE  
> disk space requirements (although instantaneous bandwidth  
> requirements aren’t all that high to stream video).  It wasn’t  
> unusual in the late 80s, early 90s to see a workstation with dozens  
> of firewire drives in a big column attached to hold the raw video.   
> Providing a suitable multi-user server architecture is quite  
> challenging.
>
>
>
> Several movies in the 90s made use of what were essentially boxes  
> of disk drives that were flown back and forth every day from one  
> location to another.  (the “nothing beats FedEx for raw bandwidth”  
> model… getting tens of Mbps network connectivity to rural Czech  
> Republic or Romania where the location shoot is happening isn’t easy.)
>
>
>
> Jim Lux
>
>
>
> From: beowulf-bounces at beowulf.org [mailto:beowulf- 
> bounces at beowulf.org] On Behalf Of Prentice Bisbal
> Sent: Monday, November 05, 2012 12:56 PM
> To: beowulf at beowulf.org
> Subject: Re: [Beowulf] Digital Image Processing via HPC/Cluster/ 
> Beowulf - Basics
>
>
>
> CJ,
>
> This article is from 14 years ago, but it might be relevant to your  
> situation.  It describes how Digital Domain used a Linux 'render  
> farm' to do the GCI for Titanic. I haven't read this article in 14  
> years, so I'm a little fuzzy on the details, but I think you might  
> learn something useful from it.
>
> http://www.linuxjournal.com/article/2494
>
> Prentice Bisbal Manager of Information Technology Rutgers Discovery  
> Informatics Institute (RDI2) Rutgers University http:// 
> rdi2.rutgers.edu
>
> On 11/03/2012 10:12 AM, CJ O'Reilly wrote:
>
> Thank you very much!
> I'll be sure to talk to the software developer about this.
> For now this project is moving slowly; still doing research (it's  
> possible simply a single powerful computer could get this work done  
> feasibly...)
> Perhaps I'll be back around in the future though!
>
> Thanks a bundle:)
>
>
>
> On Sat, Nov 3, 2012 at 9:50 PM, Lux, Jim (337C)  
> <james.p.lux at jpl.nasa.gov> wrote:
>
> Yes and no..  The application process needs to be "parallel aware",  
> but for some applications that could just mean running multiple  
> instances, one on each node, and farming the work out to them. This  
> is called "embarassingly parallel" (EP).. A good example would be  
> rendering animation frames.  Typically each frame doesn't depend on  
> the frames around it so you can just parcel the work at a frame  
> granularity to the nodes.    There are other applications which are  
> more tightly coupled and where the computation process running on  
> node N needs to know something about what's running on Node N+1 and  
> Node N-1 very frequently.   For this, applications use some sort of  
> standardized process communication library (e.g. MPI), or, perhaps  
> a library that performs a high level function (e.g. Matrix  
> inversion) that underneath uses the interprocess comm.
> 2.  Another "it depends". If the process is EP, and each node is  
> processing a different image, then your problem is one of sending  
> and retrieving images, which isn't much different from a  
> conventional file server kind of model.  If multiple processors/ 
> nodes are working on the same image, then the interconnect might be  
> more important.  It all depends on the communication  
> requirements.     Note that even EP applications can get themselves  
> fouled up in network traffic (imagine booting 1000 nodes  
> simultaneously, with them all wanting to fetch the boot image from  
> one server simultaneously)
>
>
>
>
>
> This is the place to ask..
>
>
>
>
>
> From: CJ O'Reilly <supaiku at gmail.com>
> Date: Wednesday, October 31, 2012 11:31 PM
> To: "beowulf at beowulf.org" <beowulf at beowulf.org>
> Subject: [Beowulf] Digital Image Processing via HPC/Cluster/Beowulf  
> - Basics
>
>
>
> Hello, I hope that this is a suitable place to ask this, if not, I  
> would equally appreciate some advice on where to look in lue of  
> answers to my questions:
> You may guess that I'm very new to this subject.
>
> I am currently researching the feasibility and process of  
> establishing a relatively small HPC cluster to speed up the  
> processing of large amounts of digital images.
>
> After looking at a few HPC computing software solutions listed on  
> the Wikipedia comparison of cluster software page ( http:// 
> en.wikipedia.org/wiki/Comparison_of_cluster_software ) I still have  
> only a rough understanding of how the whole system works.
>
> I have a few questions:
> 1. Do programs you wish to use via HPC platforms need to be written  
> to support HPC, and further, to support specific middleware using  
> parallel programming or something like that?
> OR
> Can you run any program on top of the HPC cluster and have it's  
> workload effectively distributed? --> How can this be done?
> 2. For something like digital image processing, where a huge amount  
> of relatively large images (14MB each) are being processed, will  
> network speed, or processing power be more of a limiting factor? Or  
> would a gigabit network suffice?
> 3. For a relatively easy HPC platform what would you recommend?
>
> Again, I hope this is an ok place to ask such a question, if not  
> please help refer me to a more suitable source.
>
>
>
>
> -- 
>
>
>
>
>
> _______________________________________________ Beowulf mailing  
> list, Beowulf at beowulf.org sponsored by Penguin Computing To change  
> your subscription (digest mode or unsubscribe) visit http:// 
> www.beowulf.org/mailman/listinfo/beowulf
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin  
> Computing
> To change your subscription (digest mode or unsubscribe) visit  
> http://www.beowulf.org/mailman/listinfo/beowulf




More information about the Beowulf mailing list