Slashdot Mirror
Virginia Tech Announces Supercomputer Plans
CousinVinnie writes "Previously noted in this Slashdot story, the administration of Virginia Tech has announced they're puchasing 1100 G5's (another story) in hopes to build a top-10 supercomputer by October 1. Tech will be spending $5.2 million over five years on the project, which should help it pull in more research money." Maybe VT can use the new computers to beef up their web site.
Not to sound like a troll, but isn't the Apple a bad machine to use for this? It's big, the fan configuration will make it extrordinarily loud, and it's built to cater to the end user, not to the embedded machine market. Yes, OSX/Darwin does work fairly well, but I'd think that the entire purpose of this computer originally would make it ill-suited to this task.
Many companies build physically smaller machines that still pack a lot of power, or sell parts to allow someone to design their own layout in a chassis. Remember, individual cases, power supplies, and the like become way overkill in such a large computer, and it would probably be cheaper to convert electricity once for a large section of the computer, supplying 12v, 5v, and 3.3v without each computer converting itself.
This just seems like the wrong way to do something thats hallmark has been in being cheap.
Do not look into laser with remaining eye.
Wait a minute - you're complaining about the cost of a G5, but go on to suggest they buy a Myrinet, a rather expensive interconnect. Something doesn't compute here.
G5: We have a PowerPc system that has been extended to use 64 bits. Your old software will run. Your new software will run faster. We have MacOSx, BSD, and Linux available, natively compiled.
There is also something to be said for the G5's parallel memory busses. It divides the ram in half, each half feeding 32 bits of the processor. You could theoretically keep your instructions on one side and data on the other, and pipeline the snot out of it.
"Learning is not compulsory... neither is survival."
--Dr.W.Edwards Deming
Was just trying to point out that even with an expensive interconnect such as Myrinet, the economics of apple just doesn't work out. But then, even if you are using a cluster of G5s, to get any reasonable super-computing power out of it, you would need a low latency (expensive) interconnect.
Opteron: Still under development.
Now tell me, on the Good/Fast/Cheap curve you design parameters lie?
"Learning is not compulsory... neither is survival."
--Dr.W.Edwards Deming
For the supercomputer to break the top five supercomputers in the world, it would have to possess 10 teraflops of memory.
I think that they mean 10teraflops of computing power, as opposed to 10terabytes of memory -- since the later would require each CPU to have 10GB of ram in it. Nonetheless, the anomaly tells me that this is a reporter not used to computer issues. (too few computer geeks at the college paper).
Free Software: Like love, it grows best when given away.
Anyone get the feeling that Apple might be pulling a Be, Inc and is trying to pull off a focus-shift?
Remember Be, the "multimedia" OS turned "Internet Appliance". Remember the death of Be. (damn, that stings. I miss the BeOS.)
Now witness Apple:
For decades, seemingly the darling of the press-production (DTP) world, catering to artists of all magnitudes, it was the computer you used to create real, bona-fide art. It attracted the freaks, the hippies, the art chicks. For many people, this was unnerving. Different people get "different" looks.
Now who's Apple targetting?
With OS X, I'm thinking geeks. We're different people, too, but in a, well, different manner. Instead of the artists, Apple's going for traditional suits, the realm of IT. It may be a matter of sheer survival that Apple penetrates here, because they don't stand a chance in these days of "homogenous" work environments.. Out with Apple (even if it works) and in with Dell WinXP machines! Linux faces the same dilemna, although Linux has some other benefits/detriments for it's widespread adoption. If Apple can show it's worth in the server room (just like Linux is doing), then maybe, just maybe, they'll start looking at Apple on the corporate desktop (just like Linux is doing).
Now, the idea of catering to suits is somewhat.. frightening. The whole damn market is different. They don't care about "look and feel", they care about numbers (see economic downturn, outsourcing to India, massive layoffs, H1B abuse, etc). This means Apple will have to change from being "cool" to utilitarian. But wait, I think I just painted myself into a corner here... Wasn't that the point of Apple? To be a tool and not an obstacle? Instead of creating computer art, we're now creating databases? Maybe Apple is on to something here...
If you were me, you'd be good lookin'. - six string samurai
to get any reasonable super-computing power out of it, you would need a low latency (expensive) interconnect.
Well, that very much depends on what type of computing you're doing. Some scientific computing is more tolerant of high-latency environments and would rather have the bandwidth.
I can't seem to find the quote from any of the articles right now, but VT is planning on using an Infiniband interconnect from Mellanox. While I don't know the relative price points, they are touting the fact that this is a high-speed interconnect that's faster than Myrinet or Quadrics at a fraction of the cost. I can't say for sure, since the Infiniband cluster we're helping to build at Stanford is not yet assembled.
This should be interesting to watch. I'll be very interested to see the $/gigaflop ratio for VT's cluster (though that doesn't have a bearing on the interconnect).
Itanium: er, um, we have a new architecture! I think RedHat has a port to it.
I admin a cluster of Itaniums. They are very fast and have not had one single hint of a problem with them.
To me, something that is completely unacceptable about the G5s for scientific use is that the machines do not support ECC memory!
My users run up to 5 days at a time across 8-10 processors, and its not cool to get a wrong answer after that run, and possibly never know about it.
I personally would not care to admin a cluster of Macs. I think they are excellent machines, but not for science.
Can someone explain the " Maybe VT can use the new computers to beef up their web site" comment? It loads perfectly fast for me. It looks pretty good. It even runs PHP, so it couldn't be a "They shouldn't use ColdFusion" type remark.
Am I missing something, or was that just a completely random comment?
________________________________________________
suwain_2
There are a bunch of people posting gripes that this was a bad idea. But I don't think it's that bad. We should at least withold judgment until we see some data. One thing's for certain, it will outperform YOUR cluster.
Among the top complaints were:
You could buy several AMD's for that.
You might be able to, but the G5's they are buying already have 2 very good processors. As long as they're dividing up tasks among processors, it's nice to have all the memory management and overhead taken care of at a level of two processors per node instead of one. To be honest, I've never seen it done before, and it could have very interesting results.
The Mac's aren't designed for this sort of thing.
We don't know all the details of this cluster because they weren't all mentioned in the story, but my hunch is that Apple might cater to them a little if they are offering to dump $5 mill on a cluster. They might package the cases differently (sans curvy plastic or with shared power supplies).
Anyway, when it comes to speed of high precision calculations, the G* chips have proven their worth. And most High Science applications fall into that range of operation. We all know that clustering and distrubuting is touchy. The cost and speed don't scale linearly. And the cost vs speed ratio definately doesn't scale literally.
There is a possibility these computer science professors know something. So we might want to see how this thing performs before we rush to judgement.
Slashdot Syndrome: the sudden, extreme urge to correct someone in order to validate one's self.
How about resale? When the projects wind down and things need upgrading, they can get maybe over half a mil' in return for offloaded desktops (or at least scrap aluminum), as opposed to 57 cents for a bunch of beige schrapnel.
You know what?
Linux IT Consulting and Domino Development in Michigan
G5: Deliverable today
Well, deliverable this weekend if you order 1,100 of them at a time; everybody else has to wait... but yeah.
$x='S24;r)>63/* h@<5+oZ)32"5cz';$me='phroggy'x$];
$x=~y+ -xz+\0-Tx+;print$_^chop$me for split'',$x;
Myrinet eats CPU when sending data?? You must be using the suckiest driver and firmware possible. And Infiniband has lower latency than Myrinet???? Infiniband is a combination of protocol and hardware and actually Infiniband has slightly higher latencies than the best MPI implementations on Myrinet. Myrinet is a just a piece of hardware. You can write firmware in Myrinet to do almost everything in the Lanai processor present in the card itself, without consuming any CPU cycles. The performance you get out of Myrinet entirely depends on the libraries you are using.
I have a sneaking suspicion that these computers aren't going to be used as a supercomputer for long. I bet they set this up, get on the supercomputer list, and then in six months or a year farm out the computers to use in computer labs around campus. Besides, I haven't heard a compelling reason why VT *needs* a supercomputer.
Mainframes have one job: to move data from point A to point B as quickly as possible, while doing a relatively minor amount of processing on the way. Mainframes are what you use when you want to process every ATM transaction that happens around the world, all at the same time. In fact, your average mainframe is not really any more powerful than a dual- or quad-CPU Intel server, raw processing wise.
Supercomputers are the exact opposite. They're stacks and stacks of CPU's that process largely independent chunks of data. They do huge amounts of processing on each chunk of data. They do *not* move data particularly well. In many cases, supercomputers are held together with Gigabit Ethernet. That's not exactly *fast*...
Different computers, different tasks.
Linux IT Consulting and Domino Development in Michigan
Apparently the PHD's at Virginia Tech disagree with you 5.2 million ($) times. Or 1,100 times, depending on your view of the world...
"Politicians find new names for institutions which under old names have become odious to the people."
Why aren't they waiting for the Xserve update?
What, there's an Xserve update?
Take a look at the heat sinks in a G5. If you can figure out how to get that into a 1U enclosure, you might want to work for Apple in the hardware design group.
-jcr
The only title of honor that a tyrant can grant is "Enemy of the State."
To get it into 25 racks, you need to get it into a 1U form factor. At that point, cooling becomes an issue - you have 40 750W power supplies *per rack* then. We're talking about a space about the size of a phone booth, and the heat equivalent of 20 hair driers on HIGH all going at once. It's gonna get TOASTY at that point. Even if you consider a 2U and expanding out to 50 racks, that's still a lot of heat per square foot.
Remember - the CPU and the memory are going to generate the same number of BTU/hour whether in a tower case or a 2U rack. And going to the *as yet unannounced* 2U just makes the cooling problem worse...
Yes, we thought a LOT about these sort of issues.
Perhaps they have entered the Reality Distortion Field (TM).
...and think about it from Jobs' perspective: This will be the biggest "Apple Computer" ever made. His ego can't handle this thing not kicking ass.
Apple's platform will probably work here, and once they prove it does, they will open a new segment of buyers, and gain plenty of positive attention. Apple's (recent) record suggests they will exceed expectations.
As to the cost...the good PR that they can derive from a working G5 supercomputer is easily worth $10m on hardware. At the end of the quarter it's hardly noticable to a company sitting on $3,500,000,000.00 in cash.
Apple gets to show MSFT they can scale OS X way past Windows. IBM gets to show Intel what it thinks of the Blue Lightning License arrangement, and Apple and IBM, by virtue of being together, get so give MSFT the finger...from way up the (top500.org) list. Not bad for their first try at this.
--
Don't forget that Pixar's always looking for high power hardware and they've got lots of new money.
Yes - you are missing that they want to have a top-notch supercomputer and they want it now, and they don't want to be fiddling with a mismatched bunch of cheap or donated equipment, and that the funding for this likely comes out of a different purse than that used for regular ongoing expenses.