[Question] _ Utilise Multiple Computers Rendering

[leducvu]

Hi everyone,

I've just applied Swarm Multiple Machine Rendering to my workflow to decrease lightmap baking time. I set up the rendering network to work. With the same scene that I'm working.

Test #1 :

Scene Stats: 700 meshes; 6,000,000 tris

- Build Time just for my main working computer is: 844 seconds

- Build Time using multiple Machine is: 365 seconds

Here are the Rendering Distribution:


So network rendering is 2,5 times faster. Nice!!! ... However,...

Test #2 :

Scene Stats: 67 meshes; 325,000 tris.

- Build Time just for my main working computer is: 1202 seconds

- Build Time using multiple Machine is: 1212 seconds

There was no difference at all!!!

Here are the Rendering Distribution:


As you can see, the first 2 PCs stop rendering really fast. Then they let the other computer to compute the rest.

Looks like this network rendering works better for complicated scene. Is there any way to make all there computer render through the end or at least to choose my main working computer (VU-PC) to be the one which render last, since it has more cores than the others.

Thank you.

[S-Dot]

I'm just guessing here but the test #2 has probably a couple of huge light maps in it that need most of the time. If you have a large mesh for an entire building for example almost everything in the scene needs to be considered for this one mesh. And this can't be spread over multiple computers. But I might be wrong.

[heartlessphil]

That's my guess too. I recently made a small test scene with 30 objects, similar lightmap for each mesh and with a 2nd computer, the build took approx 50% less time to complete. With 1 pc, it took 45 mins and with 2 pc it took 25 mins.

[darthviper107]

Each object gets assigned to a single thread, so if you have more threads than objects then it can't utilize those threads.

[goergingo]

Hey bro,

There is a topic about it.
"netfrag-sam - Why I think Lightmass is completely out of Date" --> https://forums.unrealengine.com/show...ly-out-of-Date

And i tested swarm another topic:

Yes! it is the solution for my case! i decrased "LocalJobsDefaultProcessorCount" value 38 to 12.. it worked! Thank you so much netfrag-sam!!

38-12 = 26 core for distributing.. is it too much? i tested with or without network a project.. Result is:

2 computers (12(coordinator) + 40(agent) cores) : 5.30 min
1 computer (38(coordinator) cores) : 5.43 min
what do you think?
it is nearly same result as one computer. I think it is useless.. what do you think? should be any significal differences over all progress?

and i wrote this netfrag-sam's topic:

...Lightmass calculation and quality, swarm distribution methods look old fashioned and time consuming. I tested network rendering for swarm and it is really useless. Cores does not work together. You can not use your cores efficiently . As netfrag-sam said before that one core has a job and another cores wait in vain. Unreal team must write new renderer in short term, maybe in UE5.

Unfortunately core distribution method is not beneficial. Every core has one job and another cores are waiting one core's job. They don't help each other. This is Unreal's light building method.