up vote 2 down vote favorite
share [g+] share [fb]

You know, turn a few computers on the same network into a sort of cluster to distribute the load of heavy 3D rendering, without losing their functionality as ordinary machines.

link|improve this question
3  
realtime, or for say a rendered video? – Journeyman Geek Dec 21 '11 at 3:37
feedback

closed as not constructive by random♦ Apr 2 at 20:27

This question is not a good fit to our Q&A format. We expect answers to generally involve facts, references, or specific expertise; this question will likely solicit opinion, debate, arguments, polling, or extended discussion. See the FAQ for guidance on how to improve it.

2 Answers

up vote 2 down vote

Due to the limited bandwidth and latency of a network connection, it would be quite costly (if even possible) to maintain the illusion that the cluster is actually a local graphics card. Most GPU clusters run special distributed programs that work on small work packets on their local GPUs and then aggregate their results.

The more data you need to share between nodes, the higher chance of synchronisation contention and network congestion+latency. Traditional 3D applications deposit a lot of scenes and textures in the VRAM, which in your model, would be shared among all nodes. This is quite inefficient. If you want to use a cluster, you'll need to develop your own code or adapt existing applications to other people's platforms.

link|improve this answer
2  
In otherwords, impossible. – surfasb Dec 21 '11 at 3:44
feedback
up vote 2 down vote

This is not true. Back in the day I set up 12 Pentium4 3GHz computers each with a 100Mbps NIC and 2GB Ram. I had a Gigabit NIC on a master controller all running Beowulf. I would send my videos to render to the master controller and it would compress all of the videos. I did benchmarking and the server would render the video in about 20 hours. Using my workstation, it took about 64 hours to render the video. This was all running Windows XP on my workstation and Debian Beowulf on the servers.

I eventually replaced this method with Apple computers. I was able to use FinalCut Pro and send the workloads over through a gigabit network to MacMinis. The workstation was a MacBook Pro and the workload was sent to 5 MacMinis. The cluster would render the videos in about 8 hours where my MacBook Pro would render them in about 14 hours.

link|improve this answer
1  
So... What exactly did you do? I feel this is just a big story that doesnt answer much – Simon Sheehan Dec 21 '11 at 4:16
1  
On which setup? Both configurations were able to accomplish realtime rendering of videos across a network connection. It sounds like it gives insight that his problem has a potential solution. – kobaltz Dec 21 '11 at 4:18
I like this answer. I'm curious how they hooked this up. I can take guesses, but all the guesses are labor intensive. – surfasb Dec 21 '11 at 15:48
The Debian Beowulf was really easy. I used different hardware on each P4 computer but had all going with the debian installation. I wrote installation scripts that I would execute on each one after the initial install was done. This was back in early 2000s when I did all of the Beowulf stuff. Looks like there are more recent methods. helmer.sfe.se is a cool example of one. – kobaltz Dec 21 '11 at 23:27
OSX Alternative for Cluster Rendering (circa 2008) fcproducer.com/2008/08/qmaster-cluster-buster – kobaltz Dec 21 '11 at 23:31
feedback

Not the answer you're looking for? Browse other questions tagged or ask your own question.