Cooperative rendering update

[Devin Johnston]

I've been messing around with cooperative rendering over the last week and I've learned a few things that might come in handy if you ever want to try it.

-When rendering a job you must restart the server app on each render node after every job. This must be done because the server can't get rid of the last job it rendered on its own, you must clear this job out of the computers memory or it will only render that one scene over and over again.

-The Maxwell manager can't recover servers that have encountered an error or have gone down for what ever reason. If this happens you will have to manually combing all the MXI files because the automatic process will not work.

-You can't add or remove render nodes once a job has been submitted for network rendering so make sure you are using the right computers.

-Manually merging 20 or 30 MXI files will take a very long time if the file sizes are very large. At print resolution of 3000x2550 each MXI file is 700MB and mscl -d has to load each one onto the computer you’re using before it can merge them together which can take up to 1 minute for each MXI file.

-All network folders must be accessible to all render nodes and each node needs to have that folder mapped. You also need to make sure each render node has permission to read and write data onto all other render nodes Maxwell directories.

-The image file that will be saved onto the network drive won't be the merged image from all nodes; it is only an image from one of the nodes.

-Make sure all nodes are online and ready for the job before you send it to the render farm because if one isn't ready at the time you send it there will be no way for it to pick up the job later and your automatic merger will fail.

 

That's all I can remember right now, needless to say cooperative rendering has a long way to go before it's going to be easy to use. It does work however and I've attached some examples.

 

Image 1 was what one node generated in a 9 hour period; it reached a sample level of 8 at a resolution of 3000x2550. The second image is the result of an MXI merger of about 30 or 40 nodes, I'd say the average sample level reached was probably 9 but the image is obviously much farther along than that.

[Ernest Burden III]

I assume those columns were not supposed to look like Birch trees? Is that the new noise? Not good if it is.

 

Thanks for the info.

[Devin Johnston]

No that's just a mapping mistake on my part.

[CHE]

The merged rendering looks very clean to me. What was the average render time allowed for each render node? Thanks for sharing your finding with us Devin.

[ludi]

Thank you Maxer for this comparison. Am I right, (based on my failed test's)

that V1 isn't usuable for interior Arch-vis stuff? I'm right now testing on a old quite complex interior scene with just the natural Physical Sky for lightning. But hell.....this is such bad, you woudn't recognize it as a rendering....

[skogskalle]

Image 1 was what one node generated in a 9 hour period; it reached a sample level of 8 at a resolution of 3000x2550. The second image is the result of an MXI merger of about 30 or 40 nodes, I'd say the average sample level reached was probably 9 but the image is obviously much farther along than that.

 

maybe i dont understand this correctly?... do you mean the last render was made with 30-40 nodes that each took about 10 hours to render? thats 3-400 hours of total rendertime? that cant be right... right?

[AdamT]

maybe i dont understand this correctly?... do you mean the last render was made with 30-40 nodes that each took about 10 hours to render? thats 3-400 hours of total rendertime? that cant be right... right?

Oh yeah, it can be right. The thing with M~R is, that you can have a render that's a little noisy after maybe 18-20 hours, and it looks okay--maybe useable, maybe not, depending on your criteria. But to get it completely noise free could take 200-300 hours or more.

 

This is a result of the progressive rendering system, where each level represents an improvent of roughly 50% over the previous level, but each level also takes twice as long to calculate as the previous level. So if it takes you 1 minute to reach level one:

 

L2: 2 minutes

L3: 4 minutes

L4: 8 minutes

L5: 16 minutes

L6: 32 minutes

L7: 1h 4m

L8: 2h 8m

L9: 4h 16m

L10: 8h 32m

L11: 17h 4m

L12: 34h 8m

L13: 68h 16m

L14: 136h 32m

L15: 273h 4m

 

Even though it will take 273 hours to go from L14 to L15, the difference in quality will be slight.

[Devin Johnston]

Thank you Maxer for this comparison. Am I right, (based on my failed test's) that V1 isn't usuable for interior Arch-vis stuff? I'm right now testing on a old quite complex interior scene with just the natural Physical Sky for lightning. But hell.....this is such bad, you woudn't recognize it as a rendering....

It is usable for interior and exterior work as long as you have one of these two things, lots of render time or lots of render nodes.

 

maybe i dont understand this correctly?... do you mean the last render was made with 30-40 nodes that each took about 10 hours to render? thats 3-400 hours of total rendertime? that cant be right... right?

Yes you are correct and AdamT is also correct in his explanation of how this process works. Unfortunately this is the byproduct of using the MLT way of rendering. I am hopeful that at some point NL will optimize Maxwell because I don’t think it has been done yet, there is a definite need for speed improvements.

[Fran]

Adam, it doesn't quite work that way. Sampling levels reached and in how much time doesn't follow a simple pattern. Maxwell is slow, but not that slow. And I believe that each log entry shows the total time to achieve that sample, not the time between samples.

 

From my POV, if I had 8 computers to use my 4 licenses, I'd be happy if a coop render could produce a high-res image overnight. I'm just not happy with lighting and material quality right now, so it definately isn't worth investing in 6 more boxes.

[AdamT]

Adam, it doesn't quite work that way. Sampling levels reached and in how much time doesn't follow a simple pattern. Maxwell is slow, but not that slow. And I believe that each log entry shows the total time to achieve that sample, not the time between samples.

Agreed that that was an oversimplification--it's not always exactly 2x for the next level--but in my experience it's pretty darn close to that. For a more accurate measure you should wait to get the interval between later levels, like 5 -> 6.

[Fran]

But I've reached higher sample levels in 8 or 10 hours. That doesn't mean it was a production render in that amount of time though.

[AdamT]

Oh yeah, it totally depends on the scene. One scene might be good to go at SL14 while another could require SL19 or more. Also, there's no consistency between one scene and another, e.g., one scene might reach SL 14 in 5 hours and another might take 45 hours.

 

All's I was saying is that there's this effect of roughly 2x the time for each next level, so that you might get an okay render relatively quickly, but to get all the way home could take an eternity. Or not. Depending.

[tipti]

Hi,dont know how to setup coorperative rendering..been following the instructions in the maxwell manual but still cannot set it up..need some help

[Devin Johnston]

what are you having trouble with?

[thickly]

I'm hoping someone can give me some direction on coop rendering -- I am a noob with 3d Vis so please forgive me if I don't know what I'm talking about.

 

I currently run Max/Maxwell on a P4, 3.0ghz w/HT, 1 gig RAM. I have an old PIII, 700mhz sitting around -- is it worth trying to get both of these machines running to do a coop (is that even possible?). I know setting up coop on Maxwell seems to be a bit of a pain in the a** -- so if I could get these two working together would I see a noticeable difference in my rendertimes?

 

I appreciate any insight into coop rendering at all. I know many of you are unhappy with NL/MWR -- so no need to share your greivances any further with me. Thanks in advance.

 

Sean

[Devin Johnston]

I'm hoping someone can give me some direction on coop rendering -- I am a noob with 3d Vis so please forgive me if I don't know what I'm talking about.

I currently run Max/Maxwell on a P4, 3.0ghz w/HT, 1 gig RAM. I have an old PIII, 700mhz sitting around -- is it worth trying to get both of these machines running to do a coop (is that even possible?). I know setting up coop on Maxwell seems to be a bit of a pain in the a** -- so if I could get these two working together would I see a noticeable difference in my rendertimes?

I appreciate any insight into coop rendering at all. I know many of you are unhappy with NL/MWR -- so no need to share your greivances any further with me. Thanks in advance.

Sean

 

In theory any additional computer power will make the MXI files better, the problem is that you may not be able to see any difference if your other computer can't reach a high sample level of say 10 or more.