[Maxer]

Ok so last night I had to render out one of my projects using Maxwell and cooperative rendering. I used 40 computers of different speed and let them render for 13 hours, none of them reached a sample level that was over 10. The original images were 3000x2550 and I had to merge all 40 images together manually, these are the results.

This is the result of one PC rendering for 13 hours @ 3000x2550




This is the result of 10 PC's rendering for 13 hours @ 3000x2550




This is the result of 20 PC's rendering for 13 hours @ 3000x2550




This is the result of 30 PC's rendering for 13 hours @ 3000x2550




This is the result of 40 PC's rendering for 13 hours @ 3000x2550




You can see that the biggest reduction of noise occurs between 1 & 10 nodes, after that for each additional 10 nodes the noise seems to only be reduced half as much as the time before. My conclusion is that even with 80 machines rendering all night long it is still imposable to remove all of the noise. The 40 PC's that rendered this image all had 2 Gigs of ram, and any machines that tried to render it with less than that either crashed or simply never finished.

Cooperative rendering is working although it is extremely difficult to get it up and running and producing. My IT guy has had to write a program that basically copies and renames all of the MXI files from their root folder into a network directory, while it deletes all of the MXI and MXS files. We have to run Maxwell as a service on all the render nodes and we have also had to write a service program for this to work since Maxwell can't be run as a service as it is. A third program had to be written so that the service on each machine could be shutdown and restarted since for some reason Maxwell seems to remember only the last scene it rendered and no new scenes can be added. My conclusion is that although it is working it is seriously paralyzed in many key areas which need immediate attention.

[-Adrian]

rehost the images somewhere, we're getting popups that too many users access the FTP.

[aitraaz]

Great tests Maxer - so its not quite linear, but the improvement is always visible - I'd say that 10 is the magic number, at least for us financially challenged folks

[Mihai]

I'm wondering about the hanging emitters in this scene, do you have light shining through those containers which are made of plastic or a material that is refractive? That could really add a lot to the noise, perhaps a better way is to use the ags glass and add roughness to the reflective layer. And also keep those emitters as lowpoly as possible, single poly in most cases will do fine.

Would be great if you have time to make another 13 hour test on 10 computers without the translucent plastic around the emitters. Just to see what kind of difference it would make.

[Maxer]

Mihai, the glass is a frosted AGS and the emitters are pyramids with 4 sides. To be honest I still don't know what I'm doing with this new material system so any advice is greatly appreciated. Are you saying I should replace the emitters with a single plane?

[Mihai]

If it's only 4 sides that's not too much, but the lighting from them looks like a streak, could you post a wireframe of one of the emitters together with the housing? Perhaps there's something in the geometry that contributes to the extra noise.

[Maxer]

Here is the emitter and the housing:
ftp://kirkseyviz@ftp.kirksey.com/maxwell/emitter.jpg

Also I'm trying to create a glass like this http://www.3-form.com/materials-struttu ... e_glow.php what would be the best way to do it?

[iandavis]

Maxer,

you have demonstrated clearly that for THAT scene the 'hump' is at about 10 PCs... after that not much change. So for that resolution, with that kind of scene most people could get away with their 8 node license.

However... the thing this test doesn't show is how long 20, 30, 40 PCs could get to the same quality as 10 PCs. What time incentive is there for adding an additional 10 PCs? Perhaps throw that on the BarB for 2 hours? in each case? I think to invest in a huge renderfarm the user would want to create a final image faster. the 10PC render has acceptable noise, etc... does that mean the same image quality could be gotten from 20PCs in 6 hours?



really interesting test tho maxer... thanks.

[Mihai]

The emitters look normal.....would still be interesting to see if you apply just an opaque reflective material to the cylinder.

Btw, better to post your images via http, not ftp as everybody will get 421 too many users messages when viewing this thread.

[pluMmet]

You can see that the biggest reduction of noise occurs between 1 & 10 nodes, after that for each additional 10 nodes the noise seems to only be reduced half as much as the time before.

@Mihai- I believe we had a discussion about there being a sweet spot in number of nodes As I recall I said there would be one and you said not

[aitraaz]

I mean, optimizing the scene is one issue, and its always correct to point out how to improve it, but i think that here the issue is how coop functions compared to number of nodes, so the noisier the better in this case, it really helps to see how it works. So, even if with 30 nodes its clear (thus also at 40), at 20 a bit less, at 10 noisy, etc, it makes little difference rather than at 40 nodes clear, 30 nodes less rather than at 20 nodes...The purpose is to see how coop works, its very clear - more nodes, the better, but its not linear (as would be expected) - at higher SL levels, as maxwell goes into difficulty for sampling, all nodes contribute less, because they all share that same difficulty

[tom]

Thank you for this test Maxer and good point iandavis. Of course an unbiased system will render forever even after it saturates about noise. So, you can render the same image in less hours with more rendernodes. Otherwise 10^10 computers would stuck on a nearly flatline improvement forever.

[Mihai]

You can see that the biggest reduction of noise occurs between 1 & 10 nodes, after that for each additional 10 nodes the noise seems to only be reduced half as much as the time before.

@Mihai- I believe we had a discussion about there being a sweet spot in number of nodes As I recall I said there would be one and you said not

It's the same as if you render an image on one single computer plummet......you can say there's a "sweetspot" of SL, for example there are big changes from 1 to say 12, but from that, since for each sl there are many more calculations to make, the differences will appear less and less.

So extend that to coop rendering.....what you say is it will make no difference after 10 computers, and that's simply not true.

[Maxer]

One thing I didn't post was that I let the same scene render over the weekend on a dual core dual processor machine for 63 hours, it was basically the same as the 40 node rendering.

[aitraaz]

One thing I didn't post was that I let the same scene render over the weekend on a dual core dual processor machine for 63 hours, it was basically the same as the 40 node rendering.

Ok crap don't ruin the party Maxer

It's the same as if you render an image on one single computer plummet......you can say there's a "sweetspot" of SL, for example there are big changes from 1 to say 12, but from that, since for each sl there are many more calculations to make, the differences will appear less and less.

So extend that to coop rendering.....what you say is it will make no difference after 10 computers, and that's simply not true.

Well, sure, it improves with more than ten 10, the images testify to that, but not enormously, which is logical. Its a good indication for maybe those who want to invest with a limited budget