To test this feature, visit your live site.

Edited: Jul 02, 2019

Crowdrender slower than with single box

i am continuing to try every permutation possible, but so far all my Crowdrenders have been slower, in fact, much slower than using my single main machine. i am using today's build of blender 2.80. i plan to try it in 2.79, in case the problem is to do with 2.80.

Machine specs:

master: i7 6700K + GTX 1080 in hybrid mode

node: P37 Gigabyte i7 4720HQ + GTX 970M

both machines set to CUDA.

render object: a single monkey with glass shader on a plane.

12 comments

12 Comments

doit

Jan 12, 2022

I have excactly the same issue here. The file is the demo BMW27 from blender's website (blender 3.0).

A 60 frames animation (just a camera movement) is rendered as following:

Master PC (Angry Beast): 10'24" (optix only GPU rendering)

3 nodes (Master and 2 slaves): 14'10" (slaves on cuda only GPU rendering)

In the attachement I have a comparison list of every setting I used (including load balancing) on both Master and all Slaves with their respectively times

3 nodes cannot beat the master PC with a huge different

My thoughts are, and please verify if true.

Since 1 frame is rendered in 10sec on the Master PC, the time the file needs to be distributed and loaded to each node adds an extra time that has nothing to do with the actual render time. What I mean is that if all nodes has started the rendering at the same time the final rendering time should be lower than 10sec. Now, if a scene takes 1min for 1 frame on the Master PC, probably the nodes rendering would be faster in total as the distribution time would be so small that it won't make much difference.

Of course this is a thought, I haven't test a large file but please if you can provide a solution, a tip or a verification of the above it would be great.

Thanks.

Philip

Jan 16, 2021

James, How is this rewrite going. I have a similar issue. I have four older PCs that I was intending using for helping with the rendering and a more powerful laptop as the master. What I am seeing is that the load balancing is very slowly creeping towards the correct value, but currently the slower PCs holding up the system. Each frame takes 1min 35secs to render. If however I just use the laptop on its own it takes 14secs. So a mode of one frame per computer, with the first to finish getting the next frame would suit my setup.

Details of my setup:

Blender 2.90

CR 0.2.6

Slave 1: i7-7700 16GB GTX1050 Load balancing = 0.14

Slave 2: i5-4460 8GB GTX750 Load balancing = 0.07

Slave 3: i7-4770 8GB GTX980 Load balancing = 0.36

Slave 4: i7-7700 16GB GTX1050 Load balancing = 0.16

Master: i7-10750 16GB RTX2070 Super Load balancing = 0.24

I have tried manual balancing, but still get slow rendering times.

James Crowther

Jan 19, 2021

Replying to

Hi Philip,

Sorry for the late reply, we've been swamped since the start of the year :S

Ok load balancing is going to get a whole lot better, the re-write is currently being planned out. We've got some code working, which is a new system for distributing data between nodes, we're calling that component CRESS. There will also be a new addon as well. The new addon hasn't been started but will be on our GitHub account where you can currently find free docker containers for crowdrender and blender.

So, some things you could try in the meantime:

Raise the amount of tile given to the laptop, if it is very much more powerful, use manual mode, and set the value higher until the frame render time stops going down.
Be careful about using GPUs, CR by default also uses the CPU, which can slow down some computers. You may wish to double check each node's settings before rendering and decide if you want to also use the CPU
If you do decide to use hybrid rendering, tile size will make a pretty huge difference. Often I have found that larger tile sizes can cause render time blow out as the CPU will grab tiles at the start of the render, and still be stuck rendering them long after the GPU has finished its tiles.

We'll be updating the community on the new modes for sharing out rendering work as they are coded. You'll likely see invitations going out to comment on CRESS and its API before that though since CRESS has to come first since the addon will use it to distribute data around your network.

Thanks for posting and let me know if I can help further :)

julian

Jul 03, 2019

hi james,

i completely understand regarding funding. i will be glad to do whatever i can to publicise your crowdfunding effort. we did a major crowdfunding some years ago (to try to make a feature film about the banking scandal of 2008) and we learned a few hard lessons, which i am happy to share privately (i'll send you an email).

julian

Jul 03, 2019

hi james, thanks for the reply and thanks for your add-on - it is going to be vital for my work and i am extremely keen to get it operating well.

to give you some reasonably accurate comparison aggregate animation times, i need to finish converting Jason Van Gumster's Render Time Estimation (RTE) add-on to 2.8 and ideally i will add some way of collecting all the render times in a place where i can easily compare them (eg a google spreadsheet). however, RTE works in 2.79, and i will add CR to 2.79 and try both stills and animations there too, and post all the data back here.

currently i am trying these tests on one master and two nodes, but i will be able to increase that to three nodes today, i hope.

i tried both automatic load balancing and manual. you are right - automatic gives quicker results than manual tweaking. somehow, the master machine, which is by far the most powerful, is being greatly underused and/or slowed down by CR (both i think). this is the main reason that the renders are slower with CR, i think.

i had hoped that in animation mode, CR would allow each node to render a whole frame so that as each node became free, it would simply grab the next available frame. that way the fast master machine would render more frames and over a period of frames the system would easily and automatically find the most optimal arrangement. indeed as the frame content changed, the system would again automatically adjust and whatever machine was going the fastest would get the right number of frames to suit its abilities.

one other question: can CR take advantage of hybrid rendering (only available in 2.8 i think)?

thanks,

julian

James Crowther

Jul 03, 2019

Hi Julian, how long does the scene take without CR and how long with CR?

Also, assuming you are using automatic load balancing? Automatic is generally a lot better for animation renders, you can use it for stills, but you have to be patient, rendering the scene using lower samples or resolution to begin with can help the load balancer get better at choosing the screen areas given to each computer.

If you roughly know how much faster one computer 'should' be compared to the other, then maybe try using manual mode and setting the amount of screen to render for each.