v2 daily builds - offline rendering faster than DR

[denisgo22]

With last DB from 4 2018 In addition to the problem of constantly updating IR---  I discovered for myself one more interesting problem/
Render process faster without DR !:)
After connecting for all my slaves, noise level starts to rise very quickly and practically in almost 10 times increases the rendering time, if at all it reaches the necessary parameters in approximately the time when the render was completed on one machine:)
maybe someone else has faced the same problem?

[maru]

Does it work fine with pass and time limits?

[draxdamax]

Hi

I also have some issues - in my case with the latest DB 08.01.
There is clearly a problem somewhere , at least with the noise reporting if not with the DR entirely.

I've attached a small test with/without DR.

The scene was built with DB 08.12 and the master is an I7 with slave a dual xeon E5 - V1 if it matters.

[maru]

Same question :)
Does it work fine with pass and time limits?
 

[denisgo22]

Same question :)
Does it work fine with pass and time limits?

\1.If render only on workstation with 3% noise limit- i get this 3% in 200 passes of render time/
2.If i set 200 passes and use DR /without noise limit, only passes/ with same picture i get 15% noise level in the end of render in VFB/:))
\judge for yourself:))
seems to me that using DR noise level generally does not fall below 14% never/
DB-- 8 /2018
With time option i am not testing because--i think yet everything is clear:)
-----------------------------------

[maru]

Here are my tests and conclusions:

1.7 local render with nose limit set to 5 - 75 passes, 6:56 - OK
1.7 DR with noise limit set to 5 - 88 passes, 4:15 - OK

v2 DR with noise limit set to 5 - 198+ passes, 10:34+ - NOT OK (stopped manually, reported noise level was over 20, though the image quality was already much better than in previous examples)
v2 DR with pass limit set to 75 - noise level 21, 4:17 - KIND OF OK (pretty much consistent with 1.7, except for the crazy noise level value)
v2 DR with time limit set to 4:15 - 72 passes, noise level 12,65 - KIND OF OK (pretty much consistent with 1.7, strange reported noise level again)

Update: https://forum.corona-renderer.com/index.php?topic=18802.msg118500#msg118500

[denisgo22]

Here are my tests and conclusions:

1.7 local render with nose limit set to 5 - 75 passes, 6:56 - OK
1.7 DR with noise limit set to 5 - 88 passes, 4:15 - OK

v2 DR with noise limit set to 5 - 198+ passes, 10:34+ - NOT OK (stopped manually, reported noise level was over 20, though the image quality was already much better than in previous examples)
v2 DR with pass limit set to 75 - noise level 21, 4:17 - KIND OF OK (pretty much consistent with 1.7, except for the crazy noise level value)
v2 DR with time limit set to 4:15 - 72 passes, noise level 12,65 - KIND OF OK (pretty much consistent with 1.7, strange reported noise level again)

So which conclusion?
Not correct noise report with DR render?
And quality is OK?

[TomG]

So which conclusion?
Not correct noise report with DR render?
And quality is OK?

Looking at Marcin's results, that's how it would seem to me - the rendering is ok, but the noise level reporting is incorrect with DR, giving a higher value than it should, so it continues rendering even when it has hit the needed quality, as the noise level still shows as too high.

[maru]

Yes, the conclusion is that the noise level reporting seems to be broken when using DR. Sorry I forgot to add that. :)
Anyway, this is reported directly to the dev team, so we can expect a fix soon.

Update: https://forum.corona-renderer.com/index.php?topic=18802.msg118500#msg118500

[denisgo22]

Yes, the conclusion is that the noise level reporting seems to be broken when using DR. Sorry I forgot to add that. :)
Anyway, this is reported directly to the dev team, so we can expect a fix soon.

Thanks a lot:)

[iancamarillo]

so just to make sure, this error is NOT happening in 1.7 hotfix 2 right? Thanks

[Frood]

Right, only current daily is affected.


Good Luck

[iancamarillo]

ok thx

[maru]

Guys, sorry for the confusion. It seems that my tests were not very reliable (and I did not even test v2 with no DR).
The explanation is that version 2 is using a "different unit" to report the noise level (you can see it as a different scale). This is because it is supposed to make the noise level more consistent with what was reported in Corona 1.6.

So the noise level is reported correctly, it is just using a different unit (or scale) for this.

We will of course double check whether DR is definitely working fine in the daily builds.

[PROH]

@maru - you just made me much more confused..... Could you please explain this with some noise % examples - and maybe some renders?

[maru]

The problem is that I did not do my tests properly (I did not check daily build local vs daily build DR). Maybe denisgo22 made some mistake, like having two different versions installed on the computers?
I will redo the tests ASAP and post the results here.

Sorry for the extra confusion. :)

[TomG]

Just a quick test - this was just coincidence, I was using DR to speed up making an image for the website, and noticed this.

So, same scene, set to render for 80 passes, using latest V2 daily (Jan 8th). When the render completes, here is the reported noise level:

No DR: 6.11%
DR : 20.4%

The DR was quicker (and this is a small image, with small number of passes, so DR really didn't have time to kick in), but the take away here is that the noise limit is reporting different if DR is used vs not, even if passes is used as the render limit. The noise in the images seems the same to me, so it isn't that the noise is different, just how it is being reported.

This does mean if you render to noise limit, DR images will "take longer" as they are in fact rendering to a much cleaner version of the image - as you can see if I want "6% noise" then my DR cut off would have to be set to 20% noise or so, due to the number being reported wrong.

[mike288]

Thanks for testing. It is fixed and the fix will be most probably in next daily. Track the changelog for that ...

[TomG]

Cheers Mike :)

[maru]

Tests from the newest daily build. It seems to work pretty much as expected, but I still have one question - why with DR it reported 38 passes rendered, even though pass limit was set to 30? Is it because some extra passes were sent from the node after the rendering was stopped on the master?

(30pass = pass limit set to 30)
(noise7 = noise limit set to 7)

[TomG]

That would be the cause as far as I know - the pass limit is based on the passes that the local machine has, so I believe two things can happen
1) Local machine has 29 passes, does the 30th itself, and asks the nodes for their last updates, which will add more passes and take it over 30
2) Local machine has 29 passes, and a node says "Here! Have another 6 passes!" which takes it over the 30 (plus, it will then collect any additional passes from any other nodes)

From my thinking, this will be affected by how long a pass takes to render in combination with how often the nodes send updates (fast rendering passes with longer update times, then increased chance of nodes adding a greater number of passes over the limit; longer passes and faster updates from the nodes, more likely they will just add one or two extra passes)

[mike288]

Yes exactly - master gathers all remaining passes from slaves (since they are already done) which may exceed the set limit. It just further improves the result.

[Ondra]

yep, we dont want to discard passes just so we dont overrun the limit ;)