Farm vs Single

[xscript]

Hi!
I'm buying a new comp and I have a crazy idea. Until now, I was
thinking about single workstation: dual Opteron 2.2Gh and rest to
match. Now, the part I don't like is that for a great deal of money, I
get a comp that (if it's possible to calculate this way) works at
about 4Gh. As the processing speed is my priority, that's not enough,
especially in respect to the cost.
Now, I'm probably talking nonsense, but... I always hear stories about
Render Farms, so... What if I have a Node, say 2.0Gh with 512 RAM,
motherboard, LAN and case. If I take 10 of those, I still have enough
for a Server: single low-end Xeon or even high-end P4, hard, graphics
card etc. Now I have (not that I know anything about RenderFarms) 20Gh
of processing speed. It's not just way faster than 4Gh: if a single
node is damaged, the rest od the farm is still operational.
So, I believe this post made many of you laugh. However, the question
is: am I paying everything but performance when buying a high-end
single comp, and is there more sense in buying a farm, even less
ambitious?

P.S. Sory about my english...

 

[Eli Kane]

Hi!
I'm buying a new comp and I have a crazy idea. Until now, I was
thinking about single workstation: dual Opteron 2.2Gh and rest to
match. Now, the part I don't like is that for a great deal of money, I
get a comp that (if it's possible to calculate this way) works at
about 4Gh. As the processing speed is my priority, that's not enough,
especially in respect to the cost.
Now, I'm probably talking nonsense, but... I always hear stories about
Render Farms, so... What if I have a Node, say 2.0Gh with 512 RAM,
motherboard, LAN and case. If I take 10 of those, I still have enough
for a Server: single low-end Xeon or even high-end P4, hard, graphics
card etc. Now I have (not that I know anything about RenderFarms) 20Gh
of processing speed. It's not just way faster than 4Gh: if a single
node is damaged, the rest od the farm is still operational.
So, I believe this post made many of you laugh. However, the question
is: am I paying everything but performance when buying a high-end
single comp, and is there more sense in buying a farm, even less
ambitious?

P.S. Sory about my english...

First of all, what exactly is the application you have in mind for this farm
of machines?

Second, processing power does not necessarily scale as you suggest, e.g. 5
machines at 2Ghz does not necessarily mean a 10Ghz machine. The problem you
are trying to solve and the software driving the scheduling on each machine
determines the effective boost in power you get.

Have a look at the following Beowolf FAQ for more information:
http://dune.mcs.kent.edu/~farrell/equip/beowolf/beowulf-faq.txt

Eli

 

[Wayne Stallwood]

Hi!
I'm buying a new comp and I have a crazy idea. Until now, I was
thinking about single workstation: dual Opteron 2.2Gh and rest to
match. Now, the part I don't like is that for a great deal of money, I
get a comp that (if it's possible to calculate this way) works at
about 4Gh. As the processing speed is my priority, that's not enough,
especially in respect to the cost.
Now, I'm probably talking nonsense, but... I always hear stories about
Render Farms, so... What if I have a Node, say 2.0Gh with 512 RAM,
motherboard, LAN and case. If I take 10 of those, I still have enough
for a Server: single low-end Xeon or even high-end P4, hard, graphics
card etc. Now I have (not that I know anything about RenderFarms) 20Gh
of processing speed. It's not just way faster than 4Gh: if a single
node is damaged, the rest od the farm is still operational.
So, I believe this post made many of you laugh. However, the question
is: am I paying everything but performance when buying a high-end
single comp, and is there more sense in buying a farm, even less
ambitious?

P.S. Sory about my english...

Unfortunately life is just not that simple.

A farm (or cluster) of computers is only any good for solving a particular
type of problem, that is a highly parallel one like rendering images.

Some types of problem are easily broken down into independent chunks that
can be calculated without much dependence on other bits of the puzzle being
complete. Other types of problem are more serial in nature (you have to
solve one bit to get the information required to compute the answer to the
next bit) Also with a cluster of machines, your weakest link is always the
interconnect, unless you have a very parallel problem the amount of
communication between the nodes will be quite large. You need something
with very high bandwidth and a great deal less latency than an ethernet
network.

In short you need a fair amount of knowledge, a good set of clustering tools
and some pretty specialist software to even hope of doing anything useful
with a farm of machines. However if your problem is simply rendering then
there are plenty of turn key solutions that should work out of the box
(nearly)

Even given your idea of a dual processor machine I would look carefully at
the software you are going to be using and see if it benefits from being
run on an SMP machine (not everything does)

 

[xscript]

First of all, what exactly is the application you have in mind for this farm
of machines?

Second, processing power does not necessarily scale as you suggest, e.g. 5
machines at 2Ghz does not necessarily mean a 10Ghz machine. The problem you
are trying to solve and the software driving the scheduling on each machine
determines the effective boost in power you get.

Have a look at the following Beowolf FAQ for more information:
http://dune.mcs.kent.edu/~farrell/equip/beowolf/beowulf-faq.txt

Eli

Well, at the moment I'm working in 3DsMax, but when I acquire this new
comp (or farm) I'll be working in Maya (and SoftImage).
Now, about scaleing: I understand why it doesn't scale that way.
However, rendering (especially animations) is a kind of task that can
be divided into buckets. Now if that's so, than they should be scaled
that way; each Node get's a number of buckets to render, which have no
relation to other buckets. Anyway I don't know that for sure yet
(still searching around) but it makes sense.

 

[DaveW]

Where do you plan on getting the "custom software" that's needed to run a
Render Farm?

 

[Eli Kane]

Well, at the moment I'm working in 3DsMax, but when I acquire this new
comp (or farm) I'll be working in Maya (and SoftImage).
Now, about scaleing: I understand why it doesn't scale that way.
However, rendering (especially animations) is a kind of task that can
be divided into buckets. Now if that's so, than they should be scaled
that way; each Node get's a number of buckets to render, which have no
relation to other buckets. Anyway I don't know that for sure yet
(still searching around) but it makes sense.

Well, I just took a quick peek at the Maya hardware and software
requirements on the Alias website. I did not see anything about using the
software in a cluster environment, but that does not mean it won't work the
way you are envisioning. I would suggest you contact Alias and ask them
about this since you are considering moving to this platform and are
thinking of the hardware you want to run it on. They have a list of
qualified systems (OS and hardware) that they know will work with their
software. It would probably be good to check their site for this
information before buying anything.

Since you can divide up the animation into "buckets" there is no reason that
you could not just buld a regular network of machines and farm out the
tasks on each one, then put everything together. Perhaps Maya allows you to
drive it with scripts - if this is the case, then you could easily use the
scripting laguage of your choice and control animation builds yourself over
a regular network of workstations. Note that you would probably need a
license for Maya for each machine, so you need to check with them about
this too.

Anyway, hope you get things worked out to your satisfaction.

Eli