can LightCache calculation be distributed?

[visions]

hi,
I have been wondering if there is a way to Distribute the Light Cache calculation over several Pc's? or is the LC calculation limited to the maximum number of processors or cores on the Host pc?

Is the number of passes in the LC settings area related to the number of cpu / cores? if not then what is its function?

Thanks
Vivek

[jujubee]

In current builds, the Lightcache is calculated locally. If you have multiple processors/cores on the main computer, I believe all these will be used to calculate the LC.

[DaForce]

in the latest builds LC is going to work over DR AFAIK

The number of passes is not directly locked to the CPU threads, but you will often have slightly better render times if they match up. For instance if you have it on the default 4 on a single core cpu it may be a little slower.

[DJ-Studd]

That's pretty cool DaForce. Do you know what build that kicks in?

-- DJ

[DaForce]

I dont know exactly sorry.. nor have I had time to test.
But i can say with certainty that its not until the 1.49.xx releases.

[vlado]

The light cache calculations are not distributed in DR mode, even in the latest builds.

This is because there is a great interdependency between the light cache samples (e.g. already created samples are used to improve the quality of the next samples and so on). If this process gets distributed on many machines, each machine will not know the samples from the other machines. The result is that the final light cache will be worse (more noisy) than if calculated on a single machine. This means that the more machines you use, the worse result you'll get. Of course, you can increase the number of light cache subdivs to compensate for this, but it's not very clear by how much - it may happen that you have to increase the subdivs so much that the distributed calculation is just as slow as the regular one. Further on, you will get a different result depending on how many machines were actually used. Initial versions of the light cache had this problem, but with multitheading - the result on a single-processor machine was different than that of a multi-processor one. This is where the light cache passes were introduced.

There may be a good solution for this but it would be rather complicated and is probably a topic for future research.

Best regards,
Vlado

[DaForce]

Well color me corrected.

Turns out my source was incorrect... realllllly should have checked first. Ah well.

Thanks

[dennis]

is it not possible to calc a independent 'low-res' LC on each machine and 'merge' em over the network ala PPT/maxwell

[visions]

Thanks a lot for the responses....my doubts are cleared!

It may now make sense to invest in a dual core dual opteron machine to get faster LC times or wait a year for the quad cores.

Vivek

[ene.xis]

is it not possible to calc a independent 'low-res' LC on each machine and 'merge' em over the network ala PPT/maxwell

i think u can some time ago there was a post about it...can t find it now though...

[vlado]

is it not possible to calc a independent 'low-res' LC on each machine and 'merge' em over the network ala PPT/maxwell

Technically you can - this is what is done for multithreading now. It works well for a small number of passes, but if you increase those, you will get a worse result - yu can test this yourself. Note that I mean regular light cache, not ppt mode.

Best regards,
Vlado

[serge]

is there a lightcache viewer for merging lightmaps

[DaForce]

Hey Vlado,

I dont suppose you have noticed your avatar not working... probabaly been like that for awhile.
Its linked to the old domain name.
http://www.vrayrender.com/stuff/avatars/tn_bit64.png

Might want to fix that (your avatar that is)

[ene.xis]

Note that I mean regular light cache, not ppt mode.

i also meant PPT ...

[jujubee]

Would it be possible to calculate LC into DR strips instead of sampling the whole screen at once? Perhaps in areas near overlapping, more samples are introduced to compensate for any visible bading which may occur.