Le 2022-05-04 à 14:40, Kenneth a écrit :
> I have noticed some kind of similar media-rendering behavior, especially with
> scattering media (but also with emission + absorption.)
> 
> I'm running V3.8.0 beta 1  (Windows), on a core I7 8-core/16-thread machine.
> 
> I currently use the default 'render block size' of 32X32 pixels and no 'mosaic
> preview'. The last few 'render blocks' that get processed in my images (which
> are in areas where there is media) seem to cause my machine to work *extremely*
> hard. I have noticed this in two different ways: My machine's cooling-fan speed
> greatly increases; and the Windows 'Task Manager' app shows that those last few
> blocks cause the processor to suddenly jump to nearly 100% use.
> 
> On my machine, there does not seem to be an actual slowdown in the render-- just
> some very hard processor work behind-the-scenes.
> 
> This is my *guess* as to what is going on (simply from a user's standpoint, and
> not a programmer's):
> Due to symmetric multi-processing, it would seem that each 'render block' in the
> image is being processed by a particular core/thread(?)-- or, the blocks for
> most of the render are being shared among those. But for the very last block--
> or last few blocks-- the SMP process is reducing the number of threads(?) to
> process those... in other words, maybe the last block is using only one core,
> and trying to make it do *all* of the processing... resulting in a slowdown (or
> overheating of the processor, in my case.)
> 
> Again, this is purely guesswork. I don't know the inner workings of processors
> or POV-ray's SMP stuff ;-)
> 
> 
> 
Each render block is rendered by a single core. So, your supposition is 
correct.
In cases similar to this one, reducing the render blocks MAY improve the 
performance by forcing those problematic blocks to get rendered by more 
than a single core.

So, try using +rb16 or +rb12 and see if there is any improvement.
Also, changing the rendering pattern so that the rendering start near 
the centre could help to at least hide some of the slowdown by having 
the problem blocks starting to get rendered sooner.

Post a reply to this message

Thorsten <tho### [at] trfde> wrote:
>
> If you want a better distribution of the computational load, you should
> decrease the render block size such that your object gets distributed
> between multiple render blocks.
>
> Thorsten

That works very well, thanks! From my own 'naive' analysis, I had been wondering
if a smaller block size would help, but I never tried it. (For my tests, I
forgot to mention that my media object occupies about 50% of the screen space--
unlike the OP's small object that is probably 'enclosed' by a single render
block.) Your suggestion works in both cases, of course.

A smaller render-block size has another great advantage: The scene renders
faster.
Here is my example (using scattering media):

block size of 64 pixels: 2 mins 16 sec
block size of 8 pixels:  1 min 10 sec

That's a definite improvement! And my machine's processor seems to run less hot,
based on the cooling-fan speed.

Post a reply to this message

Alain Martel <kua### [at] videotronca> wrote:
>
> So, try using +rb16 or +rb12 and see if there is any improvement.
> Also, changing the rendering pattern so that the rendering start near
> the centre could help to at least hide some of the slowdown by having
> the problem blocks starting to get rendered sooner.

Thanks Alain.

The 'extra' benefit that never occurred to me was the faster rendering time when
using a smaller block size. I thought the result might be the
opposite...probably based on my old and fuzzy memories of running media on
single-core machines, before POV-ray's SMP implementation.

Post a reply to this message

Op 05/05/2022 om 23:05 schreef Kenneth:
> Thorsten <tho### [at] trfde> wrote:
>>
>> If you want a better distribution of the computational load, you should
>> decrease the render block size such that your object gets distributed
>> between multiple render blocks.
>>
>> Thorsten
> 
> That works very well, thanks! From my own 'naive' analysis, I had been wondering
> if a smaller block size would help, but I never tried it. (For my tests, I
> forgot to mention that my media object occupies about 50% of the screen space--
> unlike the OP's small object that is probably 'enclosed' by a single render
> block.) Your suggestion works in both cases, of course.
> 
> A smaller render-block size has another great advantage: The scene renders
> faster.
> Here is my example (using scattering media):
> 
> block size of 64 pixels: 2 mins 16 sec
> block size of 8 pixels:  1 min 10 sec
> 
> That's a definite improvement! And my machine's processor seems to run less hot,
> based on the cooling-fan speed.
> 
> 
Ha! Never thought about that 'trick'!

-- 
Thomas

Post a reply to this message

Op 05/05/2022 om 19:45 schreef Alain Martel:
> So, try using +rb16 or +rb12 and see if there is any improvement.
> Also, changing the rendering pattern so that the rendering start near 
> the centre could help to at least hide some of the slowdown by having 
> the problem blocks starting to get rendered sooner.

Just nitpicking: +rb should be +bs

;-)

-- 
Thomas

Post a reply to this message

Op 5-5-2022 om 09:35 schreef Thorsten:

> If you want a better distribution of the computational load, you should 
> decrease the render block size such that your object gets distributed 
> between multiple render blocks.
> 

Is this also true when doing a stochastic render? Generally, I use:

+am3 +a0.03 +ac0.90 +r3

Is adding +bs12 (for instance) relevant in that case?

Thanks

-- 
Thomas

Post a reply to this message

Le 2022-05-06 à 02:55, Thomas de Groot a écrit :
> Op 05/05/2022 om 19:45 schreef Alain Martel:
>> So, try using +rb16 or +rb12 and see if there is any improvement.
>> Also, changing the rendering pattern so that the rendering start near 
>> the centre could help to at least hide some of the slowdown by having 
>> the problem blocks starting to get rendered sooner.
> 
> Just nitpicking: +rb should be +bs
> 
> ;-)
> 
Sorry :(

Post a reply to this message

Le 2022-05-06 à 02:00, Thomas de Groot a écrit :
> Op 05/05/2022 om 23:05 schreef Kenneth:
>> Thorsten <tho### [at] trfde> wrote:
>>>
>>> If you want a better distribution of the computational load, you should
>>> decrease the render block size such that your object gets distributed
>>> between multiple render blocks.
>>>
>>> Thorsten
>>
>> That works very well, thanks! From my own 'naive' analysis, I had been 
>> wondering
>> if a smaller block size would help, but I never tried it. (For my 
>> tests, I
>> forgot to mention that my media object occupies about 50% of the 
>> screen space--
>> unlike the OP's small object that is probably 'enclosed' by a single 
>> render
>> block.) Your suggestion works in both cases, of course.
>>
>> A smaller render-block size has another great advantage: The scene 
>> renders
>> faster.
>> Here is my example (using scattering media):
>>
>> block size of 64 pixels: 2 mins 16 sec
>> block size of 8 pixels:  1 min 10 sec
>>
>> That's a definite improvement! And my machine's processor seems to run 
>> less hot,
>> based on the cooling-fan speed.
>>
>>
> Ha! Never thought about that 'trick'!
> 
It's highly scene dependent.
On some scenes that already render fast, it can cause the renter time to 
increase.
Then, there are cases where it don't have a significant effect.

Post a reply to this message

Le 2022-05-06 à 08:11, Thomas de Groot a écrit :
> Op 5-5-2022 om 09:35 schreef Thorsten:
> 
>> If you want a better distribution of the computational load, you 
>> should decrease the render block size such that your object gets 
>> distributed between multiple render blocks.
>>
> 
> Is this also true when doing a stochastic render? Generally, I use:
> 
> +am3 +a0.03 +ac0.90 +r3
> 
> Is adding +bs12 (for instance) relevant in that case?
> 
> Thanks
> 

Yes. It is relevant in most scene where the time needed to render a 
block varies greatly.
If all blocks take about the same amount of time to render, it will have 
little to no effect.

Post a reply to this message

Op 06/05/2022 om 15:31 schreef Alain Martel:
> Le 2022-05-06 à 08:11, Thomas de Groot a écrit :
>> Op 5-5-2022 om 09:35 schreef Thorsten:
>>
>>> If you want a better distribution of the computational load, you 
>>> should decrease the render block size such that your object gets 
>>> distributed between multiple render blocks.
>>>
>>
>> Is this also true when doing a stochastic render? Generally, I use:
>>
>> +am3 +a0.03 +ac0.90 +r3
>>
>> Is adding +bs12 (for instance) relevant in that case?
>>
>> Thanks
>>
> 
> Yes. It is relevant in most scene where the time needed to render a 
> block varies greatly.
> If all blocks take about the same amount of time to render, it will have 
> little to no effect.
> 

Thanks! Good to know. I shall jot this down along with my other notes.

-- 
Thomas

Post a reply to this message