TG 2 Preview 3 (alpha) Render Times Mac Pro Dual proc quad core

Started by DeanMann, April 29, 2008, 12:46:56 pm

Previous topic - Next topic

Tangled-Universe

Quote from: Oshyan on May 01, 2008, 11:09:23 pm
With a quad core system you would never want to increase min threads over 4. There's just no good reason to. Currently you might even want to limit it to 2 for memory conservation. We're working on optimizing memory use and better utilization of quad core and higher systems should be possible in more memory-constrained situations in the future.

- Oshyan


Hi Oshyan,

I have a quadcore and I always use min. 4 threads and I run into frozen renders (render screen runs white and program does not respond).
If I understand correctly lowering the min. thread to about 2 might solve this problem while still using all 4 cores?
Thanks in advance.

Martin

Oshyan

Yes Martin, lowering the minimum to 2 might indeed help, if it's due to memory issues (as it may very well be). The next build should give you more flexibility and control over memory use, as well as performing better in multi-core systems, not to mention the addition of "large address" support for those of us on x64 systems or using the /3GB switch.

- Oshyan

Tangled-Universe

May 21, 2008, 04:41:03 am #17 Last Edit: May 21, 2008, 06:08:28 am by Tangled-Universe
Quote from: Oshyan on May 21, 2008, 01:43:28 am
Yes Martin, lowering the minimum to 2 might indeed help, if it's due to memory issues (as it may very well be). The next build should give you more flexibility and control over memory use, as well as performing better in multi-core systems, not to mention the addition of "large address" support for those of us on x64 systems or using the /3GB switch.

- Oshyan


Thanks Oshyan.

I forgot to mention, I was in a hurry to get from work :), that I have xp64, 8 GB RAM. So RAM shouldn't be that much of a problem :)

I'm currently rendering a very heavy scene with 7 VHQ, 3 HQ and 2 high-poly models @ 1200x800 with detail 1, AA 14 (grasses) and GI 2/2.
The render froze twice when using 1 instance with 4 min. threads and 16 max. threads, RAM usage was ~1.3 GB.

Yesterday I made a quick and dirty test to see the difference in adjusting the min. and max. threads.
I chose to render the default scene loaded by TG2.

- Using min. 1 and max. 2 results in 1 core of 4 rendering (25% CPU usage and ~110 MB RAM)
- Using min. 2 and max. 2 results in 3 cores, maybe 4? of 4 rendering (67% CPU usage and ~150 MB RAM)
- Using min. 2 and max. 4 results in 3 cores, maybe 4? of 4 rendering (72% CPU usage and ~190 MB RAM)
- Using min. 3 and max. 4 results in 4 cores of 4 rendering (100% CPU usage and ~220 MB RAM)
- Using min. 3 and max. 16 results in 4 cores of 4 rendering (10% CPU usage and ~260 MB RAM)
- Using min. 4 and max. 16 results in 4 cores of 4 rendering (100% CPU usage and >300 MB RAM)

The results aren't really predictable but it's obvious that even for a simple scene like this default scene the amount of RAM used increases quite much over increasing threads. Probably not to mention the difference when things are getting complicated.

Conclusively I decided to open 2 instances of TG with min. 2 and max. 2 threads and it is still rendering using ~1.3 GB RAM per instance :)

rcallicotte

Oh...JOY!!!    ;D

Quote from: Oshyan on May 21, 2008, 01:43:28 am
...not to mention the addition of "large address" support for those of us on x64 systems or using the /3GB switch.

- Oshyan
So this is Disney World.  Can we live here?

Oshyan

Martin, your system will do quite well with the next update as it enables Large Address support, but until then 8GB of RAM only helps you much if you're running multiple instances.

- Oshyan

3space

Quote from: red_planet on May 20, 2008, 10:24:34 am
Small addition...

(The initial GI pass seems to be very quick ...it's the rendering that's seems to be taking all the time.)



Hi All,

this is what I'm getting when I use 8 threads; All 8 cores peg on the GI pass. (super fast) But when the render starts all 8 cores drop to about 20-30 percent usage per core. Core usage is fairly uniform across all 8 cores.

When I switch to 3 threads the render makes full use of 3 cores for a while but then starts to drop usage of the 3 cores.

It seems the more threads I add the less efficient the use of each core is.

It seems 2 threads is the sweet spot for now.


8 core MacPro 10 GB ram

base render 2 threads 1:36s
base render 3 threads 2:06s
base render 4 threads 2:57s
base render 8 threads 5:44s


I'm also a C4D user and Cinema uses all cores wide open.

Richard

Matt

Hi Richard/3space,

Thanks for the info.

Was this using a cloud layer? If so, would you be kind enough to repeat some of those tests with the cloud layer's acceleration cache set to "None (highest detail)"? (It's found on the Quality tab). We've recently identified the acceleration cache as a potential bottleneck for multi-threading.

Other things which might be useful to disable in separate tests are the Enviro Light or the ray-traced shadows option in the renderer.

Also, the size of the subdiv cache decreases per thread as the number of threads increases. The total size (shared among all threads) can be set in the render settings.

Early multi-threading tests on quad-core machines showed very good speedup factors over single-threaded renders, so I don't know if something has changed in recent builds or whether it is much more of a problem on some machines than others.

Matt
Just because milk is white doesn't mean that clouds are made of milk.

3space

Hi Matt,

my first test where on a default scene without a cloud layer. I had the subdiv cach set to 1200 Mb.

__________________
__________________
Here are the results of my second test

base render + cloud layer "Optimal"    2 threads 2:20s
base render + cloud layer "highest"   2 threads 3:33s

base render + cloud layer "Optimal"    4 threads 4:09s
base render + cloud layer "highest"   4 threads 2:43s

base render + cloud layer "Optimal"    8 threads 4:33s
base render + cloud layer "highest"   8 threads 4:20s

__________________
__________________
Misc quick test

I noticed that the per core usage seemed more efficient when I added the clouds compared to my first renders that did not have a cloud layer; the usage of each core jump to around 40-60 percent when the cloud layer was added compared to a flat 30 percent with just the base terrain.

So I pointed the camera straight up and rendered just clouds and sky to see if the problem was just with the terrain. But just like with ground plane only, the core usage efficiency went down with just the cloud layer and stayed even throughout the rendering of the entire frame.

It seems when I include both clouds and terrain in the field of view the render uses more of the cpu resources. But if I just render the clouds or just the terrain all the cores flatline at about 30 precent each and stay at that point throughout the render.

Richard

Matt

Just because milk is white doesn't mean that clouds are made of milk.

3space

np, Let me know if you need more test. I would be glad to help.