[Radiance-general] Re: rpict memory usage keeps increasing with -aa 0

[Greg Ward]

Hi Mark,

You are correct in your deduction that Radiance loads chunks (octrees  
in this case) as they are first required during rendering, unless you  
are using the -PP option, which causes the whole scene and all  
references to be preloaded to maximize memory sharing.

Radiance doesn't do any LRU ejection of memory once it's been  
loaded.  However, virtual memory will swap out pages that are not  
accessed and this should in general be just as good.  You only need  
to make sure you have enough free disk space to accommodate your VM  
system.

I hope this helps.
-Greg

> From: Mark Stock <mstock at umich.edu>
> Date: September 2, 2008 10:48:06 AM MDT
>
> I think I can answer my own question. As the rendering progressed,  
> the memory growth slowed, and it finally ended at 6.4GB. A higher- 
> resolution rendering is showing the same fast initial growth, and  
> slow later growth, also ending near 6.4 GB
>
> I suspect Radiance loads each of the 407 individual geometry chunks  
> (octree instances?) when it is first needed. It took until near the  
> end of the render to load in the last one.
>
> Now I ask: will Radiance ever drop a chunk? If I want 24m instead  
> of 12m cylinders, I'd need ~13GB. I can only presume that Radiance  
> doesn't know this, and will continue to ask for more memory, with  
> the OS obliging with excessive page swapping.
>
> My options for larger scenes seem limited:
> 1) analyze the geometry and reduce surface overlap, or
> 2) buy a server with 16 GB RAM.
>
> Mark
>
> On Mon, 1 Sep 2008, Mark Stock wrote:
>
>> Radiance whizzes,
>>
>> I've upgraded to 8 GB of RAM, so, naturally, I found a scene that  
>> needs all that space. The rpict is eats up 4.6 GB initially, but  
>> even with "-aa 0" (turns off ambient caching, right?) I am still  
>> seeing my VIRT and RES memory sizes growing. They are at 5.3 GB  
>> after only 10 minutes of rendering. I am worried that I will not  
>> be able to render high-resolution images of this scene now. How  
>> can I prevent this?
>>
>> FYI: the scene is composed of a sky and 407 "instance"s of  
>> separately-compiled octrees, each with 20k-100k primitives.
>>
>> Mark