3 Replies Latest reply: Mar 3, 2008 6:31 PM by Warwick Teale
kaidomac Level 1 Level 1 (55 points)
I'm looking at setting up a small studio render farm, mainly for Final Cut Pro 2. I have a few questions:

1. How much does RAM matter on a render node? Assuming I go with several Quad Mac Pros, is 2gb enough? Should I slap 4 or 8 gigs in each machine?

2. Is a Gigabit network enough? Fibre Channel looks good but is out of my budget. However, I have an existing Gigabit network set up...will that be adequate?

3. How well does Qmaster actually work?

G4 Cube, Mac OS X (10.4.4)
  • Vision Level 1 Level 1 (70 points)
    GigE should work quite well.
  • tomrou Level 1 Level 1 (0 points)
    RAM: 2GB should be enough for the nodes. You will likely become CPU bound before you become RAM bound on the nodes. I have 10 machines in my cluster with 2GB of RAM and the Intel CPUs. We always hit CPU limits first.

    NETWORK: GigE should be OK. That's what we use. I see from the posts in this discussion list that FC is the fastest way to go and better because you can mount the Xsan as a "local" disk as opposed to having Qmaster create an NFS server for you.

    RELIABILITY: I get about 1,000 original clips exported to 5,000 versions before the cluster controller needs a reboot. My clips are short and sweet so your mileage may vary.
  • Warwick Teale Level 3 Level 3 (580 points)
    Hi Kaidomac, you can achieve what you want with minimal cost using QMASTER on all machines (service nodes).

    The memory you have is enough but remember that the overall benefits you will get may rely on the throughput of your slowest node. This is especially true for compressor.app but not so much for rendering setups.

    There are many posts on this forum on this but an older one at http://discussions.apple.com/message.jspa?messageID=4669434#4669434 may help.

    Typically just go and do the following:
    • go buy or borrow a GBE hub. THese are really cheap now.
    • connect all the QMASTER service nodes via this ethernet and
    • in systemprefs/networks - give them a separate subnet addressing such as,,, 1.1.1.x etc.
    • in systemprefs/qmaster: use the ETHERNET subnet only to communicate. *DO NOT MIX firewire or airport* .. it just fails for me.
    • MOUNT all the source and target file systems on ALL the service nodes. THis is simple to do manually, else surf these forums to see automated ways to do this.
    • make sure youe service nodes have the same file permissions fo rthe source and target you are using.
    • for SHAKE : just crank up the nodes in eth fileout render.. use the startup.h parameter to use QMASTER. i.e. ~/nreal/startup/include. qmaster.h= sys.useRenderQueue = "Qmaster"; also add UNC=on for the including of the host name in the path name ..make sure there are NO BLANKS in the files,path names else t wont work.
    • similar for maya etc .. look here for some ideas *http://adam.lumanation.com/408*

    For compressor.app (a transcoding 'farm"), beware that for segmented transcoding, the duration of the job will resolve to the slowest node you have. THE G5 QUADS are pretty fast for you.

    plenty of doc there.

    NO need for XSAN or FC infrastructure...