Render Manager advice needed

[tapioterava]

Hello, members of the cgarchitect forums.

 

I'm a student of a certain University of Applied Sciences in Finland, and I'm rather new to 3D modeling and rendering - I've been learning on my own for just a bit over one month, from January to February, after which I haven't had the time to do 3D. During that one month I was able to produce this using 3ds Max, Mudbox and LuxRender.

 

Now, to my request: I've been given a chance to set up a render farm on a datacenter consisting of thousands of cores, as an experiment. Me and my team will have access to a certain amount of those cores, and we are supposed build an environment suitable for rendering 3D for commercial purposes.

However, none of us have any experience in network rendering, and I'm the only one who has any experience in rendering at all, and that experience is limited to LuxRender. Therefore I ask for your advice on the matter.

 

I've been doing some research on render engines, and I've listed a few we might want to try rendering with, depending on the licensing and what the hardware in the datacenter will be (we're not entirely clear on that yet). The 3D software used in our university consists of 3ds Max, Maya and Blender.

 

1. MentalRay
   
2. IRay
   
3. VRay
   
4. Cycles
   
5. LuxRender
   
6. Maxwell Render
   
7. Octane Render
   
8. Thea Render
   
9. Indigo Render
   
10. KeyShot
    

 

I've also Googled around for different render managers, in case we might have to use those, but I haven't gone into details so far.

 

1. Backburner
   
2. Deadline
   
3. RenderPal
   
4. RebusFarm
   
5. Qube!
   
6. Dr Queue
   

 

So, for starters I'd like to ask:

 

Is a render manager necessary when setting up a render farm? And if it is, what render managers should we try, and what should we stay away from, considering ease of use since we're unexperienced. Also, are some of those render engines more preferable to others when speaking of network rendering?

 

Thank you in advance.

[cg_Butler]

It's definitely a necessity. It will manage all the jobs in the queue for you. You can prioritise, delete, amend and schedule jobs among other things.

 

I've mainly used Backburner and have got on with it on the whole for the last 15 years. It's not changed a great deal in that time and a lot of people hate it. It's options are poor in comparison to others. I've used Deadline and can vouch that it has better modes, such as being able to remotely shut down and turn machines on and off through it's interface.

 

For a manager that is essential for your usage you need to go for something stable and one that had good options for you use to setup and manage the jobs you will need to render. I'd suggest looking at Deadline. There must also be trial versions of all of them that you can download.

 

Although Backburner is what I've always used, with a bit of Deadline now and then, I'd steer away from Backburner if I was setting up a commercial render farm!

 

Sorry I can't be more help than that.

[tapioterava]

Hi, thanks for the fast reply.

 

So you're basically saying that for my situation you would choose Deadline over Backburner? Do you know if some of those render engines can't be managed with Deadline, or can it handle all of them? Can all render managers manage all render engines, or should we use different render managers for different render engines, or choose one manager that can handle most of the render engines we choose to try out?

 

Does anyone else have experience with other render managers?

[Devin Johnston]

I've used Deadline for the past two years, before that I was using Backburner and if you're going to be working on a project with thousands of nodes you'll need Deadline because Backburner will choke on it. One problem I see is the cost of implementing Deadline on a massive network like that, since they charge you per node. Since you're a university they may have a special student license that's cheaper but it's still going to be expensive.

Here's the list of supported software: http://www.thinkboxsoftware.com/deadline-5-plugins/

 

As for which engine is best to use that can depend on the project, they are all different and it will take time to learn how to efficiently use each one. For instance most of the Arch Viz industry uses either Vray or Mental Ray with Vray being the more widely used option. These engines are similar and can produce beautiful work in the right artists hands but they both have steep learning curves. Other engines like Maxwell use a brute force method of distributing light throughout a scene and can be quick to learn. They have the advantage of being able to create very realistic results right off but have the disadvantage of being very slow to render especially on interior scenes. I'm taking about being 10 times slower than Vray or MR and if time is a factor you may not be able to wait.

[tapioterava]

Hi, thanks for the reply.

 

So far based on what I've heard and read about Deadline, it seems perfect for our situation. I was thinking we might be able to make virtual machines which use several processors, thus requiring less licenses - does anyone have any experience on something like this, is it even possible? If not, then unless we can get student licenses, we might have to use a trial license, or settle for some other, cheaper render manager, possibly open source... Does anyone have experience with Dr Queue?

 

Edit: I'm somewhat familiar with the concepts behind Biased/Unbiased render engines, and based on my short experience with LuxRender, I believe it could be suitable even for animations, as long as you don't set your bar too high... in animations, especially when speaking of low budget animations, the frames don't have to be sampled to perfection, 100-500 samples/pixel should be enough (in LuxRender), since a single frame will be visible only for 24th of a second, and together the frames produce just barely visible noise into the animation.

 

On my ancient dual-core, rendering a single 720p frame even to such low S/px levels would take tens of minutes, even hours, but on my 7770HD it takes 5-15 minutes depending on the scene and the settings. Considering the computing power of the data center I'm speaking of, I think it wouldn't be an overkill to try out rendering animations with unbiased render engines.

 

However, what I was asking was if some render engines are more preferable in network rendering, in other words, do some render engines have significantly better performance when computing in parallel, or do some render engines have issues with render managers, etc.

Edited April 29, 2013 by tapioterava

[Scott Schroeder]

What OS are you using? Most large clusters like that run Linux, so scratch Max and Backburner. Find your OS then work from there.

 

When I ran render tests on the Coates Supercomputer; http://en.wikipedia.org/wiki/Coates_%28supercomputer%29 at the University where I worked while getting my advanced degree, we ran Linux and had to buy Mental Ray stand-alone. That was a pain in the butt since stand-alone at that time (I think it still is) is licensed per-CPU core and not per-machine. Plus for every copy of mental ray stand-alone you only got 10 extra network licenses. So we had mostly the nodes with 8-cores to work with, that was 8 licenses of Mental Ray for each node. It got expensive quickly as you can no doubt guess.

 

I would have rather gone with Vray but university restrictions dictated we stick with mental ray.

 

The way to render from Max over this config was to save a mental ray stand-alone file from the max render dialog and run that file over the Linux clusters. We also ran our own render manager, nothing out there was good enough for that kind of power. We based it off the Condor render manager (http://research.cs.wisc.edu/htcondor/) and added in our own functionality.

Edited April 29, 2013 by VelvetElvis

[tapioterava]

Thank you for your reply.

 

The Cluster is running Linux, hadn't even occurred to me that Backburner might not support Linux... But Max scenes can be rendered on Linux machines using other render managers, as long as the render manager runs on the OS, right? Do you think that the render managers nowadays (like Deadline, or maybe the open source Dr Queue) would be capable of handling that kind of power? I don't think we have the skill or the resources to build our own render manager. Looked up the Coates, and the cluster I'm speaking of is in the same ballpark - although I seriously doubt they will give a couple of students access to the whole cluster, we'll probably be given just a fraction of the cores to fiddle with.

[Scott Schroeder]

You'll probably be okay with deadline or other quality render manager apps. I'd stay far far far away from Backburner for this one. The thing you'll have to be aware of is that Max cannot run on Linux if you choose to use Max, so you have to get stand alone versions of the rendering software, which in turn can run on linux and open up the raw scene file. For example, mental ray you can save out a .mi file (raw scene file) for the scene and open that up on Mental Ray stand alone and fire off your jobs with that.

 

It's quite a bit of work to get this up and running. I'm not sure about your university, but mine was full of red tape and dealing with multiple departments. Then you have to deal with making sure all of your nodes have access to your central textures and what not. It's not something you can really "play with".

[tapioterava]

You'll probably be okay with deadline or other quality render manager apps. I'd stay far far far away from Backburner for this one. The thing you'll have to be aware of is that Max cannot run on Linux if you choose to use Max, so you have to get stand alone versions of the rendering software, which in turn can run on linux and open up the raw scene file. For example, mental ray you can save out a .mi file (raw scene file) for the scene and open that up on Mental Ray stand alone and fire off your jobs with that.

 

Network rendering with LuxRender is a bit similar, first you export the scene to the renderer, and then you share the .lxs file and all of the textures to all of the machines working on it, which then send their contributions back to the main host. But you can also render the same scene separately on each machine if they have the same files, and then merge the rendered .flm files together later.

 

But yeah, didn't really occur to me that since the cluster is running Linux, some programs might not work on it.. I'm kind of hoping that we could set up virtual machines that run Windows, and use several processors/machine, so we could use the programs with a bit less hassle, and wouldn't need so many licences. Do you think it would be worth it, or would it be significantly less efficient to render on virtual machines, or perhaps just not possible to accomplish?

[Scott Schroeder]

I'm not sure about the virtual machines with windows, you'd have to check with the admin people for the cluster if they even allow it. We didn't have that option, we had to work within Linux. But if you can do it, I'd at least test it out. You don't have anything to lose.

 

I think you first need to sit down with whoever runs your cluster and go over what it possible and what is not possible. That will determine your next step and what you need for software. Then, figure out if you have a budget or what you can spend. That will determine what software you can actually buy. There is a lot of initial budgetary legwork to be done before you can get to the rendering side.

 

The nice thing with using mental ray stand-alone, is both Maya and Max users can submit scenes to the main farm. They just send the .mi file and it works since it's all mental ray at the core.

[angiliaduan1]

http://3d.about.com/od/A-Guide-To-3D-Software/tp/A-List-Of-Popular-Stand-alone-Rendering-Solutions.htm

This is an article about render engine,hope can make a help.

As for render manager,you can try Fox renderfarm,use the $20 trial to test.

[Tommy L]

So what is the intended purpose? You say its an experiment to test the viability of a commercial 3d rendering farm?

If you are setting up these machines to be a commercial render farm then you'll need to think a little harder about the client. Probably the best place to start is a small cluster and getting as many applications to render with the minimal of user interaction, rather than having 'what software will scale well on Linux?' as the question.

Commercial render farms are already doing this, Im assuming its on Linux, so its possible.

[georgeursache]

hello. i have a question, don't know shure if it's related to topic. i am using vray and net render on the job, using a network of about 10 computers. but the problem is when i render i get all the resources from others and this way i can not render while they are working because they can't work well, even with low thread priority activated on my settings in vray. so far i managed to start all the computers and limit the amount of cores used by 3d max and vray spawner on the other computers, at the startup. i am looking for something similar to limit the amount of RAM used by vray so as it doesn't takes it all from the others. i want to be able to render with let's say about 50 % power of other machines during the day but at the same time the others to work well without problems related the network rendering process. Any thoughts and ideas? Does anybody managed to do this ? Is this possible ? Thanx

[Devin Johnston]

You can limit the number of CPU's used to render but I don't think you can do the same with memory because some scenes would exceed this limit and wouldn't load. Even if they did manage to fit into the memory limit they could quickly exceed it which would mean either a crash or disk caching which would further reduce the performance. Daytime rendering on workstations people are using is never a great idea, it causes way more problems than it fixes.

[Tommy L]

Buy dedicated nodes. Reducing salaried employees capacity to work is a false economy....

Typically you'll just use your compadres gHz at night time.