#R3 Network Rendering
16 messages in this thread
I've been thinking (alway dangerous) and I think I've figured out the
optimum network config for 3DSr3. If a user has 1 machine … then
with the investment of about $2.5k, they can double their rendering
performance … a very good value. Two double the performance once
again .. an investment of $5k … still a good value. Now, to
double the performance one again, would require an investment of $10k
…probably not worth it at that point. Therefore .. and Ipso-facto
(or whatever) the perfect network rendering farm for 3DSr3 is 4
machines!
>> the perfect network rendering farm for 3DSr3 is 4 machines!
Well, lots depends on what you are doing and how soon your clients
want their animations! I couldn't live with less than 10, and even
at 15 currently we are adding more!
Greg (made of money) Pyros,
>> I couldn't live with less than 10
Of course it would depend on your thruput requirements, but I was
assuming a basic system to service one animator. Its an interesting
questions though .. eh?
OK .. maybe its between 4-8 machines .. We'll ( oh boy .. here comes
that engineering degree peaking out its ugly head) what we really
have here is a systems design and analysis problem.
Assume that one animator's average task takes 3 days to design. A 3
day design job probably equates to about a 30 second animation (on
average of course) so that's appoximately 1000 frames (keep the math
easy). Also, lets assume that each frame takes 5 minuites to render
.. so that's 5000 minuites .. or about 100 hours .. or about 4
days. Now, ofcourse one would want to be able to render the
animaiton and have something to look at that same day. In the above
case that would mean having _atleast_ 20 machines .. each rendering
for 5 hours each. Now if one can't afford that level of performance
.. then the job will have to render overnight .. so we can shoot for
a 18-24 hour turn around … which means 4-5 machines.
Therefore, I was wrong, there are 2 optimum setups. The first is
greater that 20 machines which would provide you with less than 1 day
performance or a setup of 5 machines … 4 rendering machines able to
turn the average animaiton around in less than 1 day of rendering and
one machine for constant design work (and turned into a renderer in
the evenings).
.. now my brain hurts… <g>
your numbers look good to me <g>. Although, Greg does a lot of stuff
with mucho geometry in it, so your 5 minute render time would
probably need tripling. My stuff is more in the 5 minute or less
range and I get by with 3 machines, although 4 would be nice. I'm now
thinking, after many months of Lantastic etc, that a dedicated server
would be a better solution, rather than a non-dedicated server.
Rendering slaves get really unhappy when the server goes down…
>> dedicated server
That's what we've done. We took an old 386-33 and have made it our
disk server. We also have a 385-25 as an I/O server (Bernoulli, Tape
backup, etc…). We are also using Lantastic and LOVE IT! I just
purchased a 2 GB Seagate Baracuda drive and am getting ready for an
Exebyte 8mm to complete the picture. When I finally get our Matrox
Animation Xpress (MAX) system, the picture will be complete. I now
have 4 machines for rendering on the net all accessing the 2GB
drive. When the animation is complete, I will be able to take the
2GB drive (external) and hook it up to the MAX system and play the
animaion directly to the VTR … what a dream … I can't wait!
terrific! I'm green with envy! Sounds like you'll have a pretty slick
system. Many of those items you mentioned are on my wishlist (big
drive, Exabyte etc). I'd probably go with a DPS PAR, though, instead
of the Matrox since I already have a targa+. Mostly from a cost
point-of-view.
I'm going to try a Movie Express, too. Sounds like a wonderful piece
of hardware. BTW, did you go to a dedicated server because of
traffic related difficulties with R3 net rendering? I'm thinking
this may be a solution to my problems.
>> Rendering slaves get really unhappy when the server goes down…
Indeed, when working with Nwlite, I found this to be a dangerous way to work
when the server was another slave. (it's so tempting to do this when you can
work peer to peer)
I found it was best to go to a separate server for the projects data and if you
can, another for the rendered frames – specially if you're using a low cost
network like Nwlite or Lantastic.
The Lab has 2 servers (3 partitions) where one drive stores all the projects
and maps, one drive holds all the common utilities and non-changing files
(fonts, default maps, base meshes) and the last drive is the destination for
all rendered frames.
This also makes it simple for doing backups as all the important/unique data is
on the Projects server.
good points you make, Jamie. I'm sure these kind of questions are
going to crop up more and more as folk get r3 going and are looking
for optimum network configurations. I'm saving your message so I can
recite it, when those questions come up.
Thanks for the details. I've been experiencing peer-to-peer fist
fights on my little network (3 nodes) that has caused everything from
page faults to total lockups. I've moved all maps to all machines
and now things are better, but I'm still worried. I guess a cheapo
386 makes a good server, huh?
How you doing? Regarding a 386 as a file server, I'm no expert on
networks, but from what I learned investigating for ours, if your
going to be rendering a lot of frames, you should step up to an EISA
based system with an EISA SCSI2 HD and controller. The 5-10 seconds
savings per frame while writing out files adds up quick!
He Steve. I was going to go the way of the Adesk IDL and use the 386
or 486/33 as a network file server in the traditional sense, and use
the 486-on-steroids machine (EISA, SCSI II risc co-processed
controller) as the 3DS server. I'm not in a position to do that yet,
but when I get more machines I'll have to do SOMETHING.
JT: I've been experiencing peer-to-peer fist fights on my little
network (3 nodes) that has caused everything from page faults to
total lockups. I've moved all maps to all machines and now things are
better…
John, give me a call voice when you get a chance. Work (203)495-8612.
I've got 2 nodes running peer-to-peer with Lantastic AI 4.1 with no
major problems (except for a little Diaquest disk-to-tape glitch). It
sounds terribly painful to me that you need to copy your maps all
over the place. I know our hardware/software is similar, so maybe I
can help. Regarding Pharlap's page fault error, try using a good
disk surface scanning utility to check your hard drive. I found a
couple of media defects that raised heck with the swap files. Once
the bad sectors were gone, my 3DS crashes went away.
Thanks David. I'll call you, and I'll check the disk surface. Sounds
like a possibility.
JT: I've been experiencing peer-to-peer fist fights on my little
network (3 nodes) that has caused everything from page faults to
total lockups. I've moved all maps to all machines and now things are
better…
John, give me a call voice when you get a chance. Work (203)495-8612.
I've got 2 nodes running peer-to-peer with Lantastic AI 4.1 with no
major problems (except for a little Diaquest disk-to-tape glitch). It
sounds terribly painful to me that you need to copy your maps all
over the place. I know our hardware/software is similar, so maybe I
can help. Regarding Pharlap's page fault error, try using a good
disk surface scanning utility to check your hard drive. I found a
couple of media defects that raised heck with the swap files. Once
the bad sectors were gone, my 3DS crashes went away.
I would have to agree with Greg, 4 machines is not enough for anyone
running a major service but I'll agree with you that 4 certainly make
for a great 'render garden' (not quite farm size).
We currently have 16 machines on-line but 8 to 10 are used by my staff
durring the day. That leaves us short for rendering durring business
hours. I expect we'll be expanding the rendering 'field' before too
long.
Now if the 4 machines you have are Pentium CPUs, that's a whole new
ball game. My early tests with the Compaq Pentium show it to be 40%
faster than a DX-2/66. In cases like that 4 sort of = 6 and that
qualifies you for render farm status. 😉