designing a setup for DR / RT

[super gnu]

i mentioned this in another thread but im now officially intrigued.

question is, how to configure a workstation/slaves to work with DR as effectively as possible, so it comes as close as possible to the responsiveness of a single beefy machine. as it stands, with big scenes, you often have to wait half a render (plus) for the slaves to finally pick up the job and kick in.


im guessing

1) fast SSDs in each machine, so they can load max in the blink of an eye,

2) the fastest network possible (would 10gb ethernet be useful in this case, assuming the ssds are fast enough to feed it?) or would even more exotic options such as infiniband offer a benefit?

3) any software/network topology arrangements could help? im thinking of having the job folder mirrored to the c drive (ssd) of each machine so they have instant access to all resources. - or does DR package up the data on the host machine and send it over the network regardless..


4) anything else?

[Dariusz Makowski (Dadal)]

1) fast SSD if ur network can support it, also a fast RAM runningat 2000mghz+

2) Depends on ammount of nodes and SSD speeds. Also there are switches that if u configure can understand what u send and instead of asking 10x for 1 file from Master to send it to each slave they can analize data and divide it on their own sending it a lot quicker.... in a simple explanation

1 Slave > 10 gb file to >switch > 10 slave = 10 uploads 1 per each slave 1 at a time
1 Slave > 10 gb file t send > Switch > Divide on to 10 nodes> Send all data at the same time = 10x faster

3) I'm using AllwaySync or something like that, can mirror all data u need. Keep in mind u will need to map each node HDD on to ur workstation or find a better solution....

4)Not rly... how big are your scenes?

[super gnu]

the current job im on is about 4 gb of xrefs and 3-4 gb of textures...

[super gnu]

so what do i look for in these intelligent switches... and do they do stuff like this automatically or do you need to configure them manually to your specific usage scenario?

i was looking yesterday at 10gb ethernet but blimey its quite costy.. minimum 350 quid for a single port card, and small switches (6-8ports) dont seem to exist, only great big cisco and ibm jobbies for datacentres that cost 10's of thousands each.

i wonder if a workstation / server with 2-3 2port adapters could be configured to talk to 4 -6 seperate nodes over seperate links.

dammit i know almost nothing about networking.

[Dariusz Makowski (Dadal)]

Heya

Uhh nice big jobs

Inteligent switches u need to set them up as far as I was told. U have to install system on them(or they might already have one) and then tell it which IP/Machins is master, which is slave which way is data going and so on... If you in UK I can speak with my friend if he would take a job to network you up...

They are : Managed and unManaged swhitches... managed u can configure the others u dont...

As to switches, yes they do cost quite a lot... U are looking at thousands of quid and at 24+ port switch and so on....

I wonder that aswell, in windows 7 you can make "easily" bridge connection between lan ports, maybe that would save u some money...

I think in your case I would mirror my workstation to all slaves and ensure they all have that same data with the difference of render folder which you have to make a HDD if I'm not wrong so they all save to ur hdd not to their own hdds...

I had my mini network set up like that, it did work quite quick at the end...

[super gnu]

ok so ive been speaking to a networking nerd about options for connecting 3 pc's (one workstation 2 nodes) using 10gb ethernet WITHOUT a costly switch.

apparently it -is- possible using a 2 port NIC (£3-400) on the workstation, and a pair of crossover cables to the two nodes (not technically crossover cables as 10gb doesnt need them)

trouble is, he said each node would have to be on a different ip range and basically on a seperate network. they would each be able to see the workstation, but not each other.


so my question is, can i configure vray DR / backburner/deadline, to work correctly with machines on 2 seperate networks?

apologies if this is a dumb question.. i know quite a lot about workstations/watercooling/vray/3d but my lack of networking knowledge is embarrassing :P


next question i have to ask, is.. is it really worth it ?! for the price of that networking stuff i could probably get one extra node.

[Dariusz Makowski (Dadal)]

I dont think u can set up BB/DB to run on 2 networks...

I was facing the same problem a while ago like you do now. I end up buying 1 Gb switch 8 port from TPlink for 20quid and then mirroring my work on to nodes. Works smooth so far its auto mirror so I dont have to worry about anything :]