DR with progressive mode

[Fat Tony]

Hi guys,
I was wondering if anybody is having exp. on DR rendering on progressive mode. The cores of DR-ed machines are not running on full speed so i am curious if there is some tweak. It is recommended to adjust ray bundle size, but after some tests i haven't noticed any difference. thanks for input

Stefan

[Morbid Angel]

Its due to network speed. No way the DR machines can all act and transfer data over the network while utilizing 100% of the cpu. You would have to have a 10g network, perhaps to get better performance. Its one of the reasons I still use bucket for DR.

[Fat Tony]

I see,

But tbh, i don't understand why and how is that a difference. On bucket mode they still run at the same time, and they run at that situation on 100 %. Yes we don't have 10g network but still the data is not big i believe.
Would mind elaborate bit more ?

Thank you very much

[algato]

in my opinion, it would be better, the render slaves could send their render results for example, one time per minute.
Maybe with a little delay of 3-5 seconds.

[yavor.rubenov]

Hello,

When doing bucket rendering we only need to send very little information to the render machines about the bucket - the two corners of the bucket and that's it. Then the render machine renders all the needed passes - say you've set Max subdivs to 24 in a noisy area V-Ray will trace up to 576 rays for each pixel of the bucket. When it is done - it will return only the resulting colors for the bucket.

When doing progressive rendering unfortunately we can't send only some data like the buckets because every pixel has its own noise level and some pixels might already be done. So we need to send information for each pixel that needs to be rendered. You can take a look at the SampleRate render element - the white pixels are active and other colors are currently not rendered.
Then when the render machine receives those pixels it renders one pass and has to return results. And that happens many times - in the above case with Max subdivs at 24 - a render machine might need to receive pixels and return results up to 576 times.
You can control part of this with the rays per pixel setting - it controls how many rays to trace for one pixel in one single task - so higher value will mean a bit less data to send through the network. The down side of higher rays per pixel is decreased interactivity - you will see updated results less often.
Also you can increase the ray bundle size - this is like the size of buckets in bucket rendering. Again bigger size means seeing results less often.

The render slaves can't really delay sending back results because they don't render whole buckets - they render pixels in only one pass so once they are done they don't have anything more to do.

With all that said - there are still things we can do to improve utilization of render nodes in progressive mode and to decrease network traffic.

Best regards,
Yavor

[algato]

Hello Yavor,
thank you very much for the detailed explanation.

In my view there are two different focuses:

1 - inteactive mode - it's where I expect a fast reaction for testing materials or lighting parameters. It is what IPR was built for.


2 - DR with progressive - used mostly for high resolution still images, often rendered with backburner, where interactiveness is not a priority and therefore an update from DR slaves could be only per minute or two or even every a few passes (maybe some kind of parameter to control that would work well). The recalculation of the adaptiveness can take place in larger time intervals.
The highest priority is that the render nodes should render as much as possible with highest load for a long time period.

Kind regards
Alex

[algato]

One of the greatest advantages of progressive mode is, that this mode do not have the "last bucket taking forever" problem.
During progressive rendering of animation frames (without DR) I observe, that all cores are fully utilised until the frame is finished.

[yavor.rubenov]

About DR with progressive - the rays per pixel parameter is pretty much that parameter - if you set it higher - something like 8 or even 16 - you will get larger time intervals between updates and adaptivity recalculation.

[psanitra]

Hello,

Also you can increase the ray bundle size - this is like the size of buckets in bucket rendering. Again bigger size means seeing results less often.

Actually, many times exactly the opposite works for me. Reducing the ray bundle size to 64 or even 32 brings all DR nodes to 100%, specially when rendering smaller regions.

[Fat Tony]

hello,

Yes great response thank you very much for explaining. In general why i tried progressive is because this issue mentioned in other thread about for some reason different buckets, other than that there wouldn't be a reason for me personally to go to progressive.

But thank you for explanation,
Stefan