On 2/10/20 10:50 AM, Mirfaelltkeinerein wrote:
> "Mirfaelltkeinerein" <nomail@nomail> wrote:
>> "Mirfaelltkeinerein" <nomail@nomail> wrote:
>>> Hi all,
>>>
>>> I experienced a problem after upgrading a server from 1TB RAM to 2TB RAM (yes,
>>> really RAM) in our dual socket AMD EPYC machine with 48 cores/96 threads under
>>> Windows Server 2016. If I render a scene with 3500x1750 pixel, half of the cores
>>> are used (Windows processor group stuff) and the performance is quite nice. If I
>>> increase the resolution to 3600x1800 pixel (or more), it starts with 48
>>> threads/24 cores, but performance is bad and after several seconds, just one
>>> core/thread is doing all the work. Rendering can also not be stopped and runs
>>> till the end.
>>> I never experienced such a problem before. Do you have any ideas?
>>
>> Sorry, I posted the message into the wrong group. Could an Admin move it to the
>> Windows group?
>>
>> By the way, it doesn't happen, if I restrict the rendered area to the first 1000
>> rows for example, so it might be an scheduling issue with the total number of
>> pixels to be rendered.
> 
> It still does happen for larger resolutions like 3800x1900 pixel even with
> restricted render region.
> 
> 

Are you using POV-Ray v3.7 or the in progress v3.8 master branch? If the 
former, it might be worth trying the latter as Christoph fixed quite a 
few (all ?) fixed memory allocations which might cause stack issues and 
hence odd problems/crashes(1). I'm not a window user, but there are 
pre-compiled windows binaries for v3.8 releases. See:

https://github.com/POV-Ray/povray/releases

---
Aside: There was a recent TechSpot review of a dual socket 3990X system 
including POV-Ray where they mention AMD providing a POV-Ray patch to 
support all the cores/threads instead of 'just' 64. There is a pull 
request in:

https://github.com/POV-Ray/povray/pull/387

which I assume is this AMD fix, but I don't know for certain. Perhaps 
someone on the core team got email saying for sure? In any case if you 
are up to merging and compiling that pull request, it would be 
interesting to see if you can then use all your cores and threads - even 
if just on the smaller render.

When you say the render cannot be stopped and runs until the end, do you 
mean it continues with 1 thread AND still produces a good result? 
Surprised if this is what's happening, but maybe?

Why do you think your memory upgrade might be playing a part in what you 
now see? Were you able to render the 'problem' dimensions with the 
smaller amount of memory previously?

In whatever version of POV-Ray you are using have you tried the problem 
render with fewer threads using say +wt24 or something? Lastly, might 
guess with a machine this powerful, it might be running other things? If 
so, might these be causing what you see?

Bill P.

(1) - A part of this work was tangled in moving off boost threads to 
C++11 built in threads. The boost threads implementation could be 
tangled in your problem too I guess if you are not using a recent 3.8 
release.

Post a reply to this message