Why not embrace clustering?
I know I must be missing something, because it seems so obvious that Apple could really score big by embracing clustering in their pro lines.
Presently setting up a cluster is a fairly custom type job. Apple has a distinct advantage over wintel offerings in that they could support it in both hardware and in the OS so threaded aps could automatically take advantage of it.
I would think a pro line with Firewire 800 and auto-clustering would be quite attractive to graphics shops, biotech, and other raw power users. Want more power? Just buy another powermac, hook them together with firewire, and your aps run almost twice as fast (well, for highly threaded, parallelizable aps).
Such a move would be in line with the switch to dual processor configs in order to combat the mhz gap. Part of me hopes Apple continues to lag in processor speed so that they have to do something innovative like this. Then when the G5 hits, man they would scream!
Is there a good reason not to do it?
Presently setting up a cluster is a fairly custom type job. Apple has a distinct advantage over wintel offerings in that they could support it in both hardware and in the OS so threaded aps could automatically take advantage of it.
I would think a pro line with Firewire 800 and auto-clustering would be quite attractive to graphics shops, biotech, and other raw power users. Want more power? Just buy another powermac, hook them together with firewire, and your aps run almost twice as fast (well, for highly threaded, parallelizable aps).
Such a move would be in line with the switch to dual processor configs in order to combat the mhz gap. Part of me hopes Apple continues to lag in processor speed so that they have to do something innovative like this. Then when the G5 hits, man they would scream!
Is there a good reason not to do it?
Comments
Screed
[ 01-22-2003: Message edited by: giant ]</p>
Correct me if Im wrong, but I dont think I am.. Typically you dont sit in the same room as a 40++ node cluster.
[ 01-22-2003: Message edited by: blabla ]</p>
<strong>why not use gigabit ethernet like a normal person?</strong><hr></blockquote>
Oh no, 10 gigabit ethernet!
*Doing the MOSR dance*
Something along the lines of this:
Eight 1.8GHz PPC970 CPUs
64GB DDRII SDRAM (Thirty-two 2GB DIMMs)
Dual 10Gigabit Ethernet ports
NetBoot
Remote-managed
Controlled from a PowerMac workstation (w/XRAID array & dual channel Fibre Channel PCI card); which feeds & collects frames via dual 10Gigabit Ethernet connections...
Sweet!
As I understand it you can cluster computers running at different speeds, maybe Apple is worried that if they made OS level auto-clustering available then they would reduce turnover since you could milk more life out of old computers.
I doubt that, though, they have made my dual G4 faster with each OS X update (thanks quartz extreme).
<strong>In line with this, I want to see Apple put out a REAL rendernode...
Something along the lines of this:
Eight 1.8GHz PPC970 CPUs
64GB DDRII SDRAM (Thirty-two 2GB DIMMs)
Dual 10Gigabit Ethernet ports
NetBoot
Remote-managed
</strong><hr></blockquote>
You want that to be a node?
This must be for real-time rendering of an accurate-to-the-quark simulation of the Big Bang as a QuickTime VR movie, right?
As far as I'm concerned, the whole advantage to clustering is that you can get a bunch of DP machines and add them up to get something like that on a much lower budget.
As for clustering support, it's been rumored that Apple has been hard at work on kernel-level clustering for OS X Server for some time now. There really aren't any arguments against it; there are arguments in favor, namely the AppleSeed project and the eminently clusterable Xserve.
Not to even mention Rendezvous. Clusters, especially of the Beowulf variety, are a pain to set up. A clusters that built itself as you plugged in Ethernet jacks would be the be-all and end-all of Apple ease-of-use. Heck, you could even run a wireless cluster over Airport Extreme. The bandwidth would be choked somewhat, but that's not a problem for some applications. Besides, the idea of a cluster that you could add to just by dropping a computer in the room and turning it on is just too cool.
<strong>So I'm not hearing any argument against embracing clustering. C'mon, there has to be some reason.</strong><hr></blockquote>
So, you want a reason, don't you? While unquestionably possible, it may be difficult to find clients. People either are already using some clustering solution (so you have to make them switch and invest a lot of money where they've already invested) or don't really need it (so you have to convince them they do, RDF just may not prove to be enough), or cannot afford it (so you have to offer something cheap and powerful). And how many clients are there in the world? How large is the market in USD? Is it all worth the effort?
We're talking about bying a "cheap" home built AMD-based renderfarm, but it's not really what we want to do. If we could use our Macs as transparent render slaves we could also justify buing new Macs to those who really don't need one. But.. our 3D-guys just can't get enough.
Amorph, I don't like the idea of using Xserves as a renderfarm, to much waste...
I don't need USB/FireWire/Serial ports, I don't need PCI slots, I don't want a graphics connection... I don't need HDDs on every node, I just need boxes that NetBoot from a control workstation (RenderWrangler), take frame assignments from the same machine, pull them into RAM, process on a bunch of 970s, and dump the finished frame back to the RenderWrangler machine (which has a XRAID attached for all that data)...
So, CPUs, RAM, controller chips, Ethernet...
The control workstation could just as easily be the same workstation I use for Maya/Shake/FCP...
Would add a whole new definition of 'rendering in the background'...!
;^p
My friend Dean Dauger offers this product:
<a href="http://www.daugerresearch.com/" target="_blank">http://www.daugerresearch.com/</a>
it's what AppleSeed is built on.
- excerpty -
[quote] The Applied Cluster Computing Group (formerly known as the High-Performance Computing Group) at NASA's Jet Propulsion Laboratory (JPL) recently acquired 33 XServes for the purpose of using them as a parallel computing cluster. Using Pooch, provided by Dauger Research, the JPL group has begun running parallel computing code on their new XServe cluster.
<hr></blockquote>
Pooch, their clustering software, is available for download and your own clustering demo <a href="http://daugerresearch.com/pooch/whatis.html" target="_blank">here</a>
Think Zilla
For those who don't know Zilla, I'll give you a brief rundown of what it was on the NeXT/OpenSTEP machines:
Zilla is an OS-level clustering application. This allows any application to use the Zilla Framework to create and manage remote nodes [machines in a cluster], rather than home-brew their own distributed code.
With Zilla coming out in a near-future OS release [don't think 10.3... too soon] coupled with Rendezvous we will see some pretty interesting situations where all potential nodes will show themselves on a network, without having the master node manually link the remote nodes.
Couple this with Renderman 11, and you start getting some pretty cool stuff. Remember folks, the latest and greatest graphics cards are programmable. This will allow for HW accellerated Renderman.... With a few nodes and GF3 or better GPUs, we could see Toy Story in realtime.
The future's so bright, I gotta wear shades. :cool:
<strong>With higher data throughput you can doo cooler things....FW800 is daisy-chainable...</strong><hr></blockquote>
Err, but Gbit ethernet is actually 25% faster, and available on a much wider range of machines.
<strong>If we could use our Macs as transparent render slaves...</strong><hr></blockquote>
It's been a while since I've done any of that stuff in anger, but doesn't the 3D software have built-in "clustering" ability?
I'm sure that was in things like Strata years and years ago.
<strong>It's been a while since I've done any of that stuff in anger, but doesn't the 3D software have built-in "clustering" ability?</strong><hr></blockquote>They all do, but they are a pain to setup and manage. We do setup a quite large farm on weekends, but every node must have render client installed, be initialized and configured manually so it's not quite that transparent..