[PaZ]

Hi,
using several nodes in cooperative mode doesnt mean a proportional speedup, given the way samples are added.
Is it possible to code an option to segment the picture (horizontally, vertically, buckets) so each node would compute a piece of the picture with a customizable overlap ? Beside overlap margin (which may be little given unbiased nature of maxwell), this would allow rendertime to be proportionally much better scaled than now.

thanks,
Paolo

[Bubbaloo]

using several nodes in cooperative mode doesnt mean a proportional speedup, given the way samples are added.

I don't think this is a true statement. Although S.L 10 + S.L. 10 doesn't equal S.L. 20, identical cpu's in two render nodes will half the time it will take to reach a given S.L.

[PaZ]

I didnt test it, but in Fryrender this is not true, samples does not scale linearly, and out of 4 nodes you get a speedup between 2x and 3x.
I assumed this was valid in maxwell too, did you make some tests on maxwell and you saw each node scales linearly ?

thanks,
Paolo

[zdeno]

all cooperative job scales almost linear. but SL gain is almost exponential.

8 noodes would together gain 18SL in 51% time as 4 nodes would.

If You look for some time saving it is better to render many many .mxi at low sampling level e.g. 8-9 and then merge them.
There would be faked light slution without sofisticated caustic from refraction reflection and other small details. but should be clean enough.

but it is my blind guess and I could be wrong.

[PaZ]

i'll give it a try, thanks.
Anyway if you guys effectively know that nodes scales this well, my point is... pointless

thanks,
Paolo

[dmeyer]

Yes, the co-op mode is generally linear save for overheard costs such as file opening, saving, network etc.

Rendertimes are generally exponentially increasing along the SL curve. So if going from SL=1 to SL=2 requires 1 minute, going from SL=2 to SL=3 will take 2 minutes, and so on.