Adaptive rendering question

Started by Kadri, August 31, 2014, 03:05:16 PM

Previous topic - Next topic

Kadri


I don't know the real technical term for this so it might be wrong.

In the last row in an image rendering stage there are sometimes less
working buckets left for rendering that use only a fraction of the cores in the computer.
Depending on the image it might take an unnecessarily long time for that last stage.

Is it possible and are there any plans about this to distribute the idle CPU cores to those buckets?

I have an animation now where i might have lost probably 5 to 10 minutes in nearly all frames because of this.
One image takes nearly 1 hour in total and there are 300 frames.
So it makes a real difference like 1-2 days.

Because of this i use 2 instances of Terragen at the same time to use the cores at maximum all the time.
Not sure if there is any real gain i get in this way but it looks like it works...

Hetzen

It might work out that reducing the bucket size in the advanced render tab will help.

Kadri


I tried it Hetzen. But in the tests i made mostly under 128 (that is what i use as default) it doesn't get faster, even slower.

Oshyan

The reason this is difficult is because each render bucket/thread has its own cache at present. We have no mechanism for splitting a cache or sharing it, so we can't dynamically subdivide buckets. Matt has thoughts on the future for improving efficiency and this is the kind of thing he is certainly keeping in mind, but I can't say we have a specific solution for this issue coming in the near future. Fortunately in most scenes it does not make a massive difference. If you're seeing 10 minute render times to complete a single, final tile, I wonder what your overall frame render times are...

- Oshyan

Kadri


Oshyan i wrote it not very clear.
It is not that only one bucket takes 10 minutes to render.
It is about the lost time at the end that you see the working buckets going down from 8 to 1 (in my case).
It is more an average of the lost time at that stage.

I remember from the past the problem you wrote above.
I was curious if there were any news about this.
Thanks.