Network renders remarkably slow Tweet

[tangelo-otoy]

Hey all,

We run a new powerful gpu server that basically only does Octane. It has 6x 980ti and 4x 1080s in it.

When we're doing our quarter res work, we do a lot of renders in succession quickly. We have discovered through trial and error that since v3, network rendering has become much slower for us. And now that we have this new powerful render node we wanted to take full advantage of it.

However when we do one of these rendesr on a workstation it takes 20 seconds (with 1x gtx980ti) vs 11 seconds when handing off to our 10 gpu render node.

We want to make whatever changes that are necessary for that render node to run as fast as possible. Preferably we want that 20 second render to take 2 seconds. What can we do to get closer to that goal??

Thanks

[Zay]

If the scene first have to be transferred and then loaded, it will of course take time before the slave can start.
I have an orbx file 429MB in size. When I load it in my Master and hit the render button it will take 28 seconds before the slave starts to render. If I then turn the camera the slave will pick it up immediately and start to render. So it is the loading time that is the bottleneck. When the scene is already loaded on the slaves, the render should start immediately on the slaves.

I don't think you will ever get a 2 second render with loading time included when using a slave. Unless your scene is small like 1 polygon Instead build a PC with as many really powerful gpus as you can. Then you can skip the bottleneck by having the load time on the slave.

So if you need a render in a few seconds use 1 really powerful PC.
If you can wait a few minutes then a slave can be used. But be aware of the loading time.

[glimpse]

Yeah, loading & rendered information time is the issue here. To make things faster I would split render jobs & start on both machines. Lan will be quite a bottleneck for bigger snenes. Also what type of lan are You using?

[tangelo-otoy]

We have 10GB networking to everything: workstation, NAS and slave. Obviously something is getting in the way, maybe my NAS read speed, because when copying the work files to a workstation, it's a little bit faster. Who has a lot of expereince making network rendering super fast? Assuming that throwing more GPUs into the workstations isn't an option - we're not going to buy a GPu server for every desk, and the machines we have can only house two GPUs.

[Zay]

Just to make sure, you can try and test the speed from master to the slaves with this software.
http://www.totusoft.com/lanspeed1

I have a Gigabit network and the measured speed is 103MB/s to my slave.
But Octane might also have some bottlenecks. It doesn't synchronize with the slave when a network is located. And by that I mean it doesn't transfer the files at the same time as you are opening an orbx or a scene. It doesn't transfer files when you hit the render button. You have to wait for the master to load and convert the files first before it starts to transfer the files. And by that time the files can be a lot bigger. My 429MB orbx file was converted to 1.1GB when Octane started to transfer the files to the slave. But my guess is that the slave then don't need to convert the files when it receives them.