This preview has intentionally blurred sections. Sign up to view the full version.View Full Document
Unformatted text preview: Build a render farm Donquoteright0t send your old PCs to the dump: use them to build a distributed network to speed up your renders. We run through the tools and know-how youquoteright0ll need BY GRAHAM MORRISON 3D HARDWARE: THE COMPLETE GUIDE BUILD A RENDER FARM 16 T en years ago, distributed computing would have been an impossible dream for the average home users. But now, older machines and cheap beige boxes can be used to create a powerful rendering farm without either a massive fi nancial outlay or a attaining a degree in computer science. It can still be complicated, but the rewards are worth the effort. There are two widely used models for distributed computing. The fi rst has the best name: a Beowulf cluster. (If that doesnt scare you off, nothing will.) The idea is simple, even if the implementation is a little tricky. Its basically a network of computers that act like a single machine. The operating system needs to be tricked into thinking that all distributed components, including memory, storage and processing, are all parts of a single device. The operating system then manages tasks as it would on a multi-core CPU, juggling execution from one processor to another while remaining oblivious to the fact that signals that would normally be sent over minuscule bits of hardware inside a single computer are now making their way over a local network. The advantage to this approach is that any piece of software designed for a multi-CPU environment will work, as the software thinks its running on an ordinary operating system. The disadvantage is that this usually means your only choice for that operating system is Linux, and there arent many rendering applications that will work with a Linux Beowulf cluster. By far the more common variety of render farm is a normal computer cluster: a group of interconnected machines, each with its own operating system and resources. The machines just need to be told what to do, which is the job of the queue manager. Before it can start queuing jobs, the principal rendering task needs to be split into smaller separate jobs one for each computer to work on. The simplest option is to give each computer on the network a single frame of animation, so that they can be treated completely independently. Each rendering machine can then work alone on a frame. Alternatively, some queue managers support bucket-based distribution, splitting a single frame up to send different sections of the same image to different machines. The queue manager will reconstruct the frame after all the jobs have completed. Either way, the result is a dramatic improvement in total rendering times. The term quoteleft0farmquoteright0 is a surprisingly accurate description of what actually happens in a render network: processing power is farmed out to machines connected together on a local network. The machines need to be local to each other, because only a physical connection can provide the bandwidth to transfer the vast amounts of data required to render a scene. Unless youve got transfer the vast amounts of data required to render a scene....
View Full Document
- Spring '11
- Computer Graphics, Mac OS X, Render farm