Amazon EC2 Cluster GPU instance...and Cycles.

Anyone know if they play well together for a GPU render?

I’ll probably give it a go just to learn how EC2 works, but I figured I’d stop by here & see if anyone else has given it a shot & what they know.

Thanks in advance,
Todd

Well, I tinkered with the GPU instance for 4 hours or so the other day. It appears from my tinkering that Ubuntu is not the way to go & the Amazon Linux instance is. With the Ubuntu AMI, the first headache is getting the nvidia cuda stuff going from the command line with no X-display. If you can conquer that, you’re a bigger Ubuntu geek than I am :slight_smile: Without it being set up, there’s no GPU rendering. However, blender will render on the CPU just fine, just download blender & away you go.

The Amazon Linux instance is already set up with the nvidia cuda ready to go…or so they say, I didn’t get far enough to find out. If you grab blender & extract it to a folder, then run blender from the command line, it will be missing a couple of dependencies. The libGLU error can be conquered by installing mesa-libGLU with yum from the basic amazon linux repository. The libSDL error can be conquered by finding a recent RPM that has the missing file & then using YUM to install it. Voila…blender renders away on the CPU.

What I didn’t get working was GPU rendering on the Amazon Linux instance, and I know what needs to happen. In my situation, I’m not able to turn GPU rendering on because neither of my machines have a GPU. When I upload a blender file to the instance, the .blend is still set to CPU rendering, so it doesn’t use the GPU on the instance either. I haven’t conquered the python scripting to switch it after it arrives on the server.

As far as rendering on the instance with Cycles & the CPU, unpack the blend file, make all the paths relative, and render. I have a blender file that’s taking almost 1:30 per frame to render on my i7. If I send the file to my Q6600 machine also, that one takes slightly longer per frame, but I can get frames 2x as fast just by using 2 computers instead of 1. So, on my own, that file averages about :45 per frame. The cluster GPU instance renders on it’s CPU as fast as what I already have. 1 frame is :45 on the Ubuntu setup, and :43 on the Amazon Linux…so in that respect, I haven’t seen an advantage over what I’ve already got. I would imagine an 8 core mac pro would be similar.

Anyway, once I conquer the python, I’ll post another update.

todd

Well, I got the python together, it sets the GPU & CUDA correctly from what a seperate python script is telling me. However, this is frustrating: blender appears to see the GPU’s, allow you to set for the GPU & then not really USE the GPU. The CPU is pinned at 1600%, nvidia-smi is showing 0%, and it’s certainly not rendering any faster than last time.

Anyone have any ideas?

todd

I use www.renderflow.com

It has blender plugin, which makes it one click to send the entire blend file to renderflow farm, which renders Cycles.

I think it is much better than Amazon EC2 concept.

Best part is that they give you 15 Euro free credit. And What I have seen it that every render reduces your credit by only 0.15 euros

Just remember that the start frame and end frame should be set to 1 in the Render Menu of Blender, before you click “Render Flow” render button at the bottom

World has been different ever since I discovered Render Flow :slight_smile: No need to setup your own render farm or setup Amazon