Render farms are fascinate me.
I'm by no means an expert though. Far from it. Anyway.
Linux is more cost effective because you don't have any Windows licnses in your 2k budget. Each box needs a ~$100 license so with 4 machines, 20% of your budget is already gone for the software.
Now does your 2k budget include the host machine? Typical render farms are set up so you do work from one computer (like the actual editing) then send the work to the farm to be rendered.
But having a seperate host isn't really necessary. Lets say you bulid four boxes, one of them can just be the host. Hopefully Cinelerra can use the host machine to render when set to use multiple nodes.
Hardware wise, I think you need to get as much CPU power as possible. I think RAM capacity comes in a close second and hard drive throughput is third. I would priortize in that order. (all imo)
The Phenom X6 may actually be the right processor to use here. It's cheap and has a bunch of cores. Platform wise, AMD tends to also cheaper compared to Intel. I would go this route. 1055T. I would overclock it as far as it can comfortably go on stock voltage. Err on the conservative side though. Stability is paramount. (but milk stability for all it's worth!)
I'd get the stardard 4GB (2x2GB) of DDR3. Speed is relatively unimportant. I'm not sure how much RAM each node needs. But 4GB seems like alot for doing just one task. Hopefully that should be enough or close to it. The cost of upgrading to 8GB may be worth looking into. Maybe not initially though because adding an additional 4GB to the machines will be a drop in solution.
The motherboard is a place to cheap out on. Not as in buying crap, but you don't need fancy motherboards. As far a requirements, they need to support the X6, 4GB of RAM and have gigabit ethernet. That's pretty much it. Maybe a cheap 785G motherboard? One of the most important things I would keep in mind when chosing is upgradability. The ability to just drop in RAM at a later date is good. So is the ability to support newer processors. That one is harder to judge, but it's something to look into. So in the future, you won't have to get a new CPU and motherboard for each node, you just drop a new CPU in.
Each node doesn't necessarily need local storage. But each node does need to boot an OS, so I guess the simplest solution would be to give each node some kind of hard drive. All the drive will be used for is to boot the OS. The host computer needs a decently fast storage array. How fast? I'm just not sure. The capacity of this array also depends on your needs. But a Raid 0+1 array of 1TB drives seems like a decent place to start.
As far as GPUs go, I would bother with it. Cinelerra's documentation is old. When it says it can use nVidia's cards to speed up rendering, it gives an example of a 7600GS. This is pre-CUDA. So I don't think Cinelerra uses GPUs for the actual rendering. Sounds like it uses the GPU to help with playback or something. But I'm seeing little evidence that Cinelerra supports CUDA or something similar.
Power is something to consider. The easiest solution is to use 1 PSU per node. Maybe a
cheap PSU from a reliable brand.
Also an enclosure will probably be necessary. Honestly, I would just skip the enclosure and do something like this.
(dammit can't find the link I'm looking for. Basically you take four really long (2-3ft) screws and you use nuts to stack the motherboards on top of each other. Makes a nice tower.) But if an enclosure is necessary, I would cheap out on these. Rackmount seems wayyy out of budget. Get some cheap mATX cases. You could save even more money by using a case that comes with a PSU. The built in PSU is crappy for sure, but budget is important and there's little risk of putting other components at risk. I would
THINK that if one of these PSUs fails it won't destroy anything.
Everything I've said are my comments and opinions that I've developed from prior knowledge as well as a bit of reading through Cinelerra's documentation. Don't take what I say as final. They're all just suggestions. Hope this helps.
Interesting Links: