Distributed Rendering wish...

[fooprobe]

Hi - I'm using DR to render to several machines in the office, and I notice that after the GI pass, the rendering waits for the last CPU bucket to finish its pass before the remaining CPUs move on to the rendering pass. All too often it's the slowest machine working on the hardest bucket of the scene, and all the rest of the CPUs are idle, it's a real bottleneck. Is there any way to move the CPUs forward to the next pass as soon as they finish, without witing for all the CPUs to finish the first pass?

Thanks!

[dalomar_cg]

Re: Distributed Rendering wish...

First off, this is a Chaos Group issue since its completely tied into the render process itself. Secondly, the answer is most likely no. Its more or less a memory corruption issue. If we try and read and write to memory at the same time that tends to cause corruption, and if we proceeded to the next phase of the render we couldn't guarantee that when that last bucket needs to get written to memory that we wouldn't be trying to read at the same time. Plus, I would imagine that the changes in the render process that would be required to achieve this would be likely be pretty significant.

The only thing that I can really suggest on this one would be to use Single Bucket mode if its really causing that much of an issue. Single Bucket mode calculates the IR map and the image all at the same time, so any waiting would be for the last buckets of the final image. Try it out and see if that helps.

[Micha_cg]

Re: Distributed Rendering wish...

An other solution would be, if the bucket order options would work for the IM pass too. Than, the order could be choosed so, that the heavy buckets are started first. I have seen it often, that a different bucket order would help to speedup the IM pass. Damien, will it be possible per VfR or should I ask Vlado (because it's an engine issue)? Your answer for my request was so, that it sounds like you are not sure it's possible.

The "bucket" mode cause visible bucket artefacts, also if I disable DR. How can this artefacts avoided?

[dalomar_cg]

Re: Distributed Rendering wish...

There is a reason why bucket order does not work with IR passes...triangulation is the most efficient mode in terms of geometry being loaded/unloaded per bucket. Since the bucket order is considered more of a convenience option than something that has a practical function, supporting a different order for IR passes isn't very useful. You can ask Vlado, but since he probably wouldn't have even put in the different bucket sequences if people hadn't bugged him for it, I don't think that he'd be that interested in changing the IM mode too.

Bucket mode isn't a very good solution, but it may be worth at least checking out in some cases. Maybe its an issue of finding out that the grass really is greener...

[Micha_cg]

Re: Distributed Rendering wish...

There is a reason why bucket order does not work with IR passes...triangulation is the most efficient mode in terms of geometry being loaded/unloaded per bucket. Since the bucket order is considered more of a convenience option than something that has a practical function, supporting a different order for IR passes isn't very useful.

I will post a link to this discussion at the chaosgroup forum. To make it complete, here the quote from the email conversation some months befor:

mainly, as there's no real significant advantage of one over the other.

Hi Damien,
there is a big advantage: if an bucket needs 30 seconds, because a complicated blury refraction must be calculated, than it is better, it is started at the beginning of a rendering calculation, because in this time the other buckets can render the rest of image (for example in 30 seconds also). If the expensive bucket is started at the end, than all other buckets are ready and only the last buckets needs 30 seconds extra (total 1 min instead approx. 30 seconds). This is a simple example. Some times I wait for two or three buckets more than one minute to get the calculation done and all other buckets are ready. Do you see the advantage? Most important for fast calculated images like for an animation - 30 seconds more or less can be a speed increase of 50%..100%. If you render 1000 frames, than this can be some hours.

[fooprobe]

Re: Distributed Rendering wish...

maybe I don't understand "Single Bucket Mode" - if I set Region Division width and height to 1 and "Means" to "Region Count" then I only get one bucket, but then I'm not using DR at all (not even using more than one processor!) and the render is much slower. did you mean something else?

unfortunately, messing with the bucket order isn't specific enough to be very useful... hard buckets could be anywhere, they don't occur in nice geometric patterns.

thanks for the glimpse into how complicated this question really is, I appreciate the serious explanation.

[Micha_cg]

Re: Distributed Rendering wish...

"Single Bucket Mode" means the "bucket" mode at the IM read cache options.

unfortunately, messing with the bucket order isn't specific enough to be very useful... hard buckets could be anywhere, they don't occur in nice geometric patterns.

Right, but some times you know where they are placed. For example if I render a product image than it could be, that at the left upper corner is placed a transparent frozen plastic with liquid inside. In this case I would set the bucket order to "left-right" so that the calcualtion starts at the difficult part of the image. Also it could be helpful if the bucket mouse tracking could be used for the IM pass. So, the user could set the mouse at the difficult area.