Dynamic Splitting is ineffective

[isoparmesan]

One of the major reasons I was excited to upgrade to V-Ray 3.1 from 2.4 was the new dynamic splitting feature. However it seems to be falling a bit short of my expectations.

The way I had imagined it working was that as a frame is getting rendered, if there are buckets still being rendered by the time there are spare cores, the regions still being processed would be split up dynamically amongst whatever cores were now available.

However, this is not what happens. Instead it just "dynamically" splits up the last few regions of an image, even if there are buckets that are still processing (and these are always the ones that increase inefficiency because there many cores not being utilized).

I suppose this works if you have the scene is equally dense all the way across, but usually what happens to me is there's one object that has a particular reflection that needs a lot of samples, or maybe an object that is motion blurred. This is likely to be somewhere closer to the center of the image and doesn't count as the "last few areas of the image" due to its placement, even though it actually IS the last part to finish rendering.

[vlado]

Yes, there's definitely room for improvement.

Best regards,
Vlado

[isoparmesan]

So I'm not using it wrong?

[super gnu]

no, it works as you described. i believe its actually very hard for vray to do what it should ideally, which is split the hard-to-render buckets rather than the last ones.

issue is, if a bucket is stuck, vray will already have rendered samples all over the bucket. so to dynamically split it, vray would have to throw away all the calculations done so far, and re-render the whole bucket with smaller ones. how does it know when to do this? after a minute? an hour? how small to make the new buckets? maybe the difficult bit is only in one corner of the original bucket, so youd have the same issue again.. does it wait another X minutes before throwing the new stuck bucket away and starting again?


id imagine it could do some kind of optimisation based on the time buckets took in a previous render, but then it would have to know if the camera/ scene had changed...

im sure the Chaos guys have some ideas, but its not a simple thing to do.

[isoparmesan]

I would think that it could look at the average render time per bucket for a frame, and then have a threshold option to abandon and split up buckets that are x amount over that threshold when there are cores available. I don't know if there's a percentage completion on a per bucket basis, but that could also be a factor.

Actually, it wouldn't even necessarily have to abandon the bucket. It could just use whatever available cores to split up a duplicate of the remaining bucket, and then whoever finishes first wins.

It's just frustrating when I have 23 available cores and my render hangs on one bucket for most of the render time.

[Neilg]

You might get more milage out of it by changing the bucket order to best reflect where the detail is. you've got top/bottom, left/right and spiral (using reverse as appropriate) to hit the hard bits first.

[isoparmesan]

Yeah, I've been trying to do that. Hard to have a blanket setting that fits all frames of an animation. This issue is made much worse by motion blur. I'll have an in blurred background that renders quickly and one object that moves that takes forever, even though it's only a small fraction of the frame

[t.petrov]

There are some things we have in mind how to improve the situation, but we've not tried them yet.