Processor Bricks
Maccentral has an <a href="http://maccentral.macworld.com/news/0212/13.ipfirewire.php"; target="_blank">article</a> today talking about the rerelease of IP over Firewire. I know this is not fresh news, but I would like to use it illustrate a thought I have been having, namely processor bricks.
We all know that clustering is possible, people have been doing for a long time. But it is not an easy thing to set up for the average user. By combining Rendezvous' dynamic IP discovery, IP over Firewire, Firewire 1 or 2, and Macs with two firwire ports, you have an ability to daisy chain devices through firewire and have them instantly talk to each other.
So imagine a relatively small tower-like design that housed nothing but a small motherboard with firewire built in, ram, and a processor(s). It could sit on the floor next to a tower. Tech heads correct me if I am missing something, but that could be all you would need for in essence a "blade server". As soon as you plug this "brick" into your main computer, it is identified and added to your processing power.
I believe that there was an "xGrid" thread earlier that covered very well the possibilities of plug and play clustering. But when combined with IP over firewire you can see a very cool possibility of being able to easily expand your processing power to fit your needs. OSX in already capable of this, and the hardware exists. So it is only really a matter of time I hope...
couple of questions..
does the firewire bus have the ability to provide enough power to such devices, eliminating the need for power chords?
am i missing anything on the hardware side needed for clustering?
We all know that clustering is possible, people have been doing for a long time. But it is not an easy thing to set up for the average user. By combining Rendezvous' dynamic IP discovery, IP over Firewire, Firewire 1 or 2, and Macs with two firwire ports, you have an ability to daisy chain devices through firewire and have them instantly talk to each other.
So imagine a relatively small tower-like design that housed nothing but a small motherboard with firewire built in, ram, and a processor(s). It could sit on the floor next to a tower. Tech heads correct me if I am missing something, but that could be all you would need for in essence a "blade server". As soon as you plug this "brick" into your main computer, it is identified and added to your processing power.
I believe that there was an "xGrid" thread earlier that covered very well the possibilities of plug and play clustering. But when combined with IP over firewire you can see a very cool possibility of being able to easily expand your processing power to fit your needs. OSX in already capable of this, and the hardware exists. So it is only really a matter of time I hope...
couple of questions..
does the firewire bus have the ability to provide enough power to such devices, eliminating the need for power chords?
am i missing anything on the hardware side needed for clustering?
Comments
<strong>IP over Firewire, Firewire 1 or 2, and Macs with two firwire ports, you have an ability to daisy chain devices through firewire and have them instantly talk to each other.</strong><hr></blockquote>
Can I just point out something here, which seems to be an issue that everyone is missing. Chaining two Firewire enabled machines together gives you a great 400Mb between them, daisy chaining more than two together quickly erodes the total speed obtainable because the "network" is actually reliant on its weakest link, which is any given Firewire port. ie with five machines, and four of them trying to send data to one (the one on the end), they're sharing 400Mb so are getting 100Mb each.
Fact is 1Gb ethernet is 2.5 times faster anyway.
Additionally, if you actually start having a reasonable sized cluster, say 12-16 machines, a 100Mb network with decent switches is probably going to be better anyway.
So no silver bullet here.
I always though it would be more interesting if the different modules snapped together using fibreoptics, though... for communication
And no power... each item uses solar cells to convert light from an intense light source to power needed for each module.
Of course, this was just a pipe dream... No doubt it will always remain such... Comunications this way could never be as fast as light since it has to be constantly converted back and forth as data and light.
Just rambling, I guess... carry on... nothing to see here.
A couple of things. As far as not being needed because of the Xserve, you missed my point entirely. Of course a rack of servers would do the same thing, but for the average user they only have a desktop. I am talking about the ability to enhance the processing power of desktop machines easily. Obviously this would make clustering servers easier as well, but I am talking about the average user or independent professional (think photoshop).
Now as far as the limitations of firewire, if I recall correctly Firewire 2 is scalable up to well over a Gigabit per sec when using fibre. Firewire 2 is a protocol that can be utilized over many different types of wiring, like Cat5, fibre, and current Firewire lines. (please jump in if i am wrong). My point being is that Firewire in any form is plug and play, much easier to hook up. Thus you have grid clustering capabilities in an easy environment. BTW, not all clusters are based on gigabit ethernet.
It may not be a brilliant idea, but I don't think it is a dumb one.
<strong>Wish I could give ya chuckle. But no, that was not me.
A couple of things. As far as not being needed because of the Xserve, you missed my point entirely. Of course a rack of servers would do the same thing, but for the average user they only have a desktop. I am talking about the ability to enhance the processing power of desktop machines easily. Obviously this would make clustering servers easier as well, but I am talking about the average user or independent professional (think photoshop).
Now as far as the limitations of firewire, if I recall correctly Firewire 2 is scalable up to well over a Gigabit per sec when using fibre. Firewire 2 is a protocol that can be utilized over many different types of wiring, like Cat5, fibre, and current Firewire lines. (please jump in if i am wrong). My point being is that Firewire in any form is plug and play, much easier to hook up. Thus you have grid clustering capabilities in an easy environment. BTW, not all clusters are based on gigabit ethernet.
It may not be a brilliant idea, but I don't think it is a dumb one.</strong><hr></blockquote>
Your correct, Firewire 3 is scalable to 3.2 Ghz over stranded glass fiber according to the 1394b specs. I think that your idea is good to a degree, however I think that Apple, or any other hardware manufacturer will want you to buy full scale computers to use this tech. They will probably some type of avenue for Xserve becouse it is a high end system, so its customers will demand it. Look at it as Apples answer to the Blade server. For home use, it will be a way to hook into your VCR, DVD, Sterio and TV, not to build a personal cluster, that is unless you buy a bunch of their regular computers, or Xserve's. If Apple were worried about the "Pro's" having a second machine at home, then they would sell a less expensive PM.
By the way, I think that your idea is a good one, I also think that Apple should come out with an expandable Home Server that you could just pop in a second, third, ...16th processor in when you have the money, or better yet a $999 computer without a monitor, but I dont think that any of these will be put out by Apple in the next 12 months...
<strong>Advanced Firewire switches/routers for clustering would be nice to add too.</strong><hr></blockquote>
That's the only chance it has of working, as noted above.
<strong>Well, the whole point of rendevouz is to make tcp/ip truly plug 'n play, isn't it? So why would firewire be better for the average user?</strong><hr></blockquote>
You're right in a sense, but firewire is still easier than ethernet for most people. No crossover cables, no hubs.
And yes, that Yellow Dog brick is very similar to what I had in mind.
This is precisely a market where such technology would be useless. Using Photoshop as an example, the only instance where you'd need the processing power of ten macs is on a huge, multigigabyte image file. In such an instance, firewire would be a far huger bottleneck than just leaving the whole image on one machine and processing it there. And moreover, it's not like throwing more cpus at a 'slowness' problem is the answer. Many, many processes done in photoshop are a single thread. A single thread runs the same on one processor machines as on 128 processor machines. Because it only uses one processor. You need highly, HIGHLY threaded software made specifically for ditributed computing. Such applications are not found for home markets, but for nuclear bomb detonation simulations and car crash simulations, which involve minimal data and massive processing.
Third, there is an overhead for ditributed computing, and overhead that is far greater than you imagine. Two macs connected together w/ firewire would be maybe 20% faster than one Mac alone. Adding more Macs would diminish that number further to a point where 8 Macs would actually be slower than one.
Clustering works for Google.com, not iTunes MP3 encoding.
<strong>Are you the same guy that posted about having building block computers? The premise was something like this: create a fast common bus that you could attach all sorts of things to. They would just snap together like legos. Need more ram? Just snap some on. Need more cpu, just snap another one on. I think that idea was posted twice by the same guy in like 3 months. I got a chuckle out of it.</strong><hr></blockquote>
That might have been me. Nobody responded to my post, so I posted it again. Guess people aren't much for wild larks. Though I've seen plenty of them around here. My idea involved a 2048 bit bus. hehe <img src="graemlins/lol.gif" border="0" alt="[Laughing]" /> If your gonna dream... Dream big.
CPUs can be paired with memory. Individual threads run on their own CPUs, stored in memory. If a single brick contained memory and 1/2/4/X CPUs, I don't see why you would need such an absurd bus.
Barto