Hardware Setup for DR ( Ethernet, Optical, Direct ?)

[Delineate]

Hey Guys hopefully someone can shine some knowledge on this.

We have a growing render farm;

The problem we're having is the time it takes for the other nodes to start rendering (DR). How can we speed that up?

Currently we are using the built in Lan port on the motherboard running through a router/switch to the render slaves.

Is there a quicker way to connect to the nodes? Faster Ethernet? Optics? some form of Direct Connection? Can you please recommend some hardware we could use. Thanks in advance

[arobbert]

Speeding up your networks and server may help. I've found distributing large IR and light cache maps slows things down a lot. To speed this up you can move them to local drives on each node. See this link: http://www.chaosgroup.com/forums/vbu...ick-in-so-long

I feel it's better to have a few really fast nodes compared to many slower ones since it reduces the bandwidth requirements of your network and server. It also reduces the number of nodes to copy the maps to which takes time.

[jujubee]

There's a few main issues which occur:

1) File server. If you're using a consumer level one, you might get 10-12 mbps. Windows server 2008 can get around 50 mbps. High end dedicated file servers can hit up above 60-80 mbps.
2) Hard Drives. You can have the fastest network in the world but you're still going to be limited by hard drive speeds. Some SATA drives have a jumper with 2.0 and 3.0 settings. Look for that.
3) Switches. Consumer level switches are not designed to handle a large simultaneous load/connection. You'll need something business class. Dell makes good pro switches.
4) Teaming. Some motherboards have dual ethernet ports on the back. Some support teaming for doubling up your connection. Might be worth looking into.
5) Jumbo Frames support. Check to see if your cards support Jumbo Frames. If there's a machine on your farm that doesn't, then weird things can happen so watch out.

A few more tips:

1) Don't render on the machine that's serving the files.
2) If you can, process IR and LC maps on one machine then drop it after it's finished on the server where all the machines can see them. The biggest slow down I find is having to force the slaves to calculate them. See above.
3) Even better, copy down textures, etc. LOCALLY to each slave in a directory they can all access. No matter how fast and convenient the network, this direct access still is the most reliable method. See above.

[Delineate]

thanks for the tips jujubee, very comprehensive. Will check the hardware and look for the bottleneck. Thanks again