PDA

View Full Version : LW9.6 Screamnet/LWSN - Speed tests



SandroBenigno
04-13-2009, 05:23 PM
I'm testing the LWSN to render a scene. I'm using two machines, one with XP and the other with Linux. I'm not understanding the speed results. Rendering directly inside LW9.6 is fastest than two machines together. Can Somebody explain it?

I registered the render time of frame 78. Take a look:

Direct render inside LW9.6 (XP machine) -> 101.3 sec

(LWSN on -2 mode)
Network render LW + LWSN (both on XP machine) -> 256.8 sec
Network render LW(XP machine) + LWSN (Linux machine) -> 272 sec

(LWSN on -3 mode)
LWSN (XP machine) -> 260 sec
D-storm LWSN Controller + LWSN (both on XP machine) -> 260 sec

With that results... rendering directly inside LW is better than LWSN with two machines!!! Why????? Can you explain it?

Paddie
04-16-2009, 03:27 AM
With that results... rendering directly inside LW is better than LWSN with two machines!!! Why????? Can you explain it?

You canīt render ONE Frame on multiple Machines. Every LWSN Node renders only complete Frames. If the Node has finished one Frame he gets the next Frame from the "Master"

For Splitframe rendering I only know ButterflyNetRender

HTH

Paddie

Hieron
04-16-2009, 05:11 AM
I assume you are not trying split frame but an animation right?

Wild guess, do you have cached GI on, with the update to auto/always?
If so, the cached GI is constantly added to so frames along the road render way more GI samples.

But no clue, not much info :)

SandroBenigno
04-16-2009, 06:45 AM
You canīt render ONE Frame on multiple Machines. Every LWSN Node renders only complete Frames.

Paddie, I understand your point, but I'm not trying to split. I send the same frame to the nodes at different times to performance test. Interesting... I didn't knew that Butterfly are capable to split frame to render it in multiple nodes. That's great!


I assume you are not trying split frame but an animation right?
Wild guess, do you have cached GI on, with the update to auto/always?
If so, the cached GI is constantly added to so frames along the road render way more GI samples.
But no clue, not much info :)

Hieron, I don't baked radiosity (static or animated) and the radiosity cache is disabled. I know that cached radiosity tends to create more sample points at the render time and spend more time to finish... but it's disabled. So... what's happen? It's crazy.

Maybe it work if I enable chache radiosity and set it to locked... I will try and post here the results.

Thank you for reply, guys!

Hugs!

SandroBenigno
04-16-2009, 09:38 AM
I discovered where was the problem. The difference was on the two LW9.cfg files.

R:\LW_config\LW9.cfg (segment memory: 32MB) default
R:\SN_config\LW9.cfg (segment memory: 1024MB) changed

This particular scenes setup works better with LOW segment memory(strange but makes sense). With 2 segments occurs the trick: the first half calcs all the radiosity thing. On the second half, all recalculations are jumped, so the time is optimized. Firstly I think it occurs cause the frame is empty on lower space. But I tested on frames that uses all the space and... IT WORKS TO!!!

http://www.newtek.com/forums/attachment.php?attachmentid=72478&stc=1&d=1239895858

Experiment yourself and post the results here... it's strange, but works. :beerchug:

Hieron
04-16-2009, 01:56 PM
That makes no sense to me, bizarre :)


ps: I use Butterfly netrender, it has a great way of updating from your current config files. So never a mismatch in config file versions.. Big thumbs up for them and the easy usage.

biliousfrog
04-17-2009, 01:27 AM
You didn't list the specs of the machines that you're using...perhaps the networked machines don't have enough memory to allow rendering in a single segment.

I was always advised to put the segment memory limit as high as possible without having a detrimental effect on render times. If the scene is too complex and you don't have the resources, rendering to a single segment can actually increase render times. 1024mb would be very high for a 32bit system, 256mb would be what I'd expect. I actually have mine set to 512mb for my 64bit computers with 4gb-8gb RAM which still allows a single segment on most scenes up to A4 300dpi.

What I suspect is that you've just determined that your segment memory limit was too high.

SandroBenigno
04-17-2009, 05:13 AM
You didn't list the specs of the machines that you're using...perhaps the networked machines don't have enough memory to allow rendering in a single segment.


Biliousfrog, I discovered that isn't a LWSN thing. So the final tests uses only the SN config (lw9.cfg) on a local machine. This machine have more RAM than the nodes. My mistake was put the segment memory to high... but the same occurs on machines with lot of RAM.


I was always advised to put the segment memory limit as high as possible without having a detrimental effect on render times. If the scene is too complex and you don't have the resources, rendering to a single segment can actually increase render times. 1024mb would be very high for a 32bit system, 256mb would be what I'd expect. I actually have mine set to 512mb for my 64bit computers with 4gb-8gb RAM which still allows a single segment on most scenes up to A4 300dpi.

What I suspect is that you've just determined that your segment memory limit was too high.

I understand your point. But, it's not the case. The machine of the screenshoot above have a XP32Bit with 2G of RAM. It's a lower config... but, like you saw on screenshoot, I tried 256MB and 64MB too, with the same results. The speed ups only when the two segments occurs, with 32MB.

I tried now with XP64 with 4G... same results (it's bizarre... I now that) :screwy:

Thank's for reply, Biliousfrog. :D

SandroBenigno
04-17-2009, 10:32 AM
You can test it yourself. You will see that behavior when rendering a frame of the attached scene with memory segments 256MB and 32MB. Don't needs to render on LWSN nodes... you will see on your local LW9.6 installation. With memory segment of 32MB your render will speeds up. You can disable Raytrace Reflection and Raytrace Shadows to test more quickly.

It strange behavior is caused by motion blur and radiosity together. So, in any way, every time I will test that thing, before render this kind of scene.

It's crazy... Is it a bug? Is it a new trick? I don't know. But it works.

pixym
04-18-2009, 05:08 AM
This is very very weird that two segments make render faster than one...