Slashdot Mirror
Time For A Cray Comeback?
Boone^ writes "The New York Times has an article (free reg. req.) talking about Cray Inc.'s recent resurgence in the realm of supercomputing. It discusses a bit of Cray's decline when the Cold War ended, "the occupation" under SGI, and the rebirth of the company after the Tera (now Cray Inc.) purchase. Recently Cray Inc. has been shipping their vector-based Cray X1 machine, designing ASCI Red Storm, and recently was one of 3 (also Sun, IBM) to win a large DARPA contract (PDF link) to design and develop a PetaFlops machine by 2010. Could Cray Inc. be poised for a comeback? Wall Street seems to think so."
Partner Link
Posting as Anonymous Coward, please award my Karma to starving children in the world.
Naturally. We have another Bush in the Whitehouse, and I even hear the Wang Chung is making a comeback -- so why not Cray?
Roving Web-Teleoperated Robot
SCO vs. Cray
Many scientists are very concern about state of supercomputing in US. Hopefully new generation of supercomputers improve this situation.
Of course I expect that...in my Playstation IV,
equipped with an opto-quantic Emotion Engine VI
and a couple petabytes of holographic storage.
-><- no
There are still MANY applications for supercomputers. A lot of people think that linux/beo-clusters are going to be replacing supercomputers of the Cray/NEC/IBM variant. Not true. There are still many research, scientific, and military applications that require machines developed not for "slow" distributed number crunching, but require ultra high speed processor and memory architechtures.
So definately, time for Cray to come back and retake the supercomputer industry crown.
memory bandwidth
It's SUPER! Off-the-shelf components are just kind of "Meh."
There's a whole bunch of PETAFlops outside of McDonalds right now having a sit in and screaming about how fur is murder.
I had to literally step on their faces to get a Big Mac.
I don't need no instructions to know how to rock!!!!
If you look at the list of top 100 supercomputers, there are systems that are almost 15 years old or even older (not sure on a few). I know these take years to build and are multibillion dollar projects, but between time has got to be a killer.
Then there's the question of ... what do you need a supercomputer for? The applications are pretty limited for a need for a petaflop computer, unless your doing mass storage, cryptography (cracking), or simulations.
Don't get me wrong I'm all about nuclear testing being done in 1's and 0's instead of in the ocean or in the desert, but how big of a bomb do you really need when it's estimated theres enough nukes to blast the entire land surface of the earth 3 times over.
Ignore the "p2p is theft" trolls, they're just uninformed
Bandwidth.
Well, a well engineered supercomputer has much less overhead than a cluster. One superfast processor doesn't have to deal with interprocessor communcations like a cluster does.
And if your supercomputer has multiple processors, they are generally made to cooperate nicely to speed efficiency. Whereas a cluster has to go through ethernet and hardware layers to communicate between nodes. Granted that is fast, but on-board communication is faster.
It seems strange, but a multiple processor computer can actually perform a task slower than just one processor working on the problem if the program and os aren't designed well. So a lot of the value of a supercomputer comes in its design, and the reputation of the manufacturer. And Cray is pretty reliable in my book.
But the REAL key to the potential comeback of the Cray computer will be whether or not it still has cool bubbles! Wow!!! Cray computing... the inventor of case mods.
Slashdot Syndrome: the sudden, extreme urge to correct someone in order to validate one's self.
Cray died. Anything else is just bartering on his name.
can someone explain to me what the benefit of a moving van is compared to buying a fleet of pintos?
Are they gonna code BCD (binary coded decimal) in their "supercomputers"? I wouldn't call any computer that does anything for the sake of decimal "super", including conversions. Stop acknowledging decimal!
-Libertarian secular transhumanist
Saw this earlier today. My first thought was how cool it was to see the old cray logo again. More than that, 'tho, I can see some real possibilities here. Since home computers are increasingly looking like supercomputers of yore, it will be interesting to see if any of this technology trickles down to the home market. I want a CRAY AMD box.
Memory to processor feeding: std ots processors are often idle because the memory subsystem cannot feed the processor fast enough. This is bad now. It will be getting a lot worse.
Interconnections between processors: this goes beyond merely processors on a board, but between boxes. The bus architectures out there for the std ots hardware get saturated very quickly. This gets worse between boxes. In addition the latency on Myranet and Quadrics (compared to what Cray et al do) is horrible even if it is excellent compared to ethernet.
Problem set vs architecture: Not all problems map out well to clusters, or even SMP boxen. Some map best to vector machines. Some map best to tightly integrated MPPs. Some map out to moderately tight clusters. Some are just plain 'embarassingly parallel'. Others are highly threaded and don't work well on vector or scalar machines. etc, etc. The architecture ought to match the problem set.
MTBF: Mean time between failures. Commodity hardware goes kaputt much more often. A cluster capable of teraflop performance of custom hardware tends to need constant and evil levels of care and feeding: ie you better have a grad student on roller blades.
Those are just off the top of my head. I am sure that others will Tell you others before I can post again. ;)
Summarized: bandwidth, latency, problem set, and failure rate.
HTH.
Do you know why the road less traveled by is littered with the bones of the unwary?
Didn't Sun basically buy out or hire away a bunch of Cray, Inc.? I always heard the E10000 was actually a Cray product. Oh, and just to brag, I have a blue jacket with a picture of a Y-MP-90 on the back with the words, "CRAY - WORLD'S FASTEST SUPERCOMPUTERS". Too cool for words. Ebay rules.
Shutting down free speech with violence isn't fighting fascism. It IS fascism!
SCO vs Nike
Look at me, I'm a stock analyst!
I know nothing about how supercomputers actually work, but here's a hypothetical comparrison in terms of mhz:
A. 1 supercomputer running @ 100,000 mhz
B. 100 nodes running @ 1000 mhz
Naturally choice B. would be cheaper to assemble and operate, but A has the advantage of actually processing all that information locally at those speeds, which would be better for things that require real time processing.
Games, for example, require real time processing, but of course no one runs games on supercomputers. I dunno what exactly they run, but there must be plenty of applications where local real-time performance is more desireable than a large group of cheap nodes.
The unofficial
I was just on their site looking at the machines. Weird, weird, weird.
Mmmm......sacrelicious.
Not Found
:)
The requested URL / was not found on this server.
Additionally, a 404 Not Found error was encountered while trying to use an ErrorDocument to handle the request.
Apache/1.3.26 Server at www.tera.com Port 80
---
Looks like they need to host their website on a SUPERcomputer to handle a Slashdotting! (Noooooobody expects a Slashdotting!
They run SETI@Home. You could finish a unit in about 30 seconds on one of those things...
503 Sig Unavailable
The Signature could not be accessed. Please try again later or contact the administrator
The home page at Cray for the Cascade project.
There are some interesting PDFs there. Chew, mull, and consider.
Also consider what Horst Simon, head of NERSC said here too.
Do you know why the road less traveled by is littered with the bones of the unwary?
Maybe now that there's once again a major player in the computer market with machine casing designs even SILLIER than Apple's, the rest of the Geek Community will give us a little slack..
But I still want an old one in my basement for a couch.
"Sic Semper Tyrannosaurus Rex."
Second, (yes, I work for Cray so now I'm going to put in a sales pitch
Finally, there's memory. Lots of it. A single system image supercomputer can have terabytes of memory in one kernel image. You're simply not going to get that in a single PC cabinet.
Finally, in case anyone doubts that vectors, big memory, and large bandwidth can make a good system, the fastest machine in the world right now is the Japanese "Earth Simulator" machine which is an NEC SX machine. That is somewhat similar in architecture to a Cray in that it has large bandwidth and vectors.
Go Badgers! -- #include "std/disclaimer.h"
somebody.. please stop this new simoniker before it's too late, i have some work to do, really :)
- Arwen, I'm your father, Agent Smith.
- Well, you're just Smith, but my father is Aerosmith!
Hahahaha. Have you ever actually run a supercomputer? They tend to have much higher failure rates then normal servers. Couple of reasons: first, they push the envelope of a given technology. The sweet spot for stability is not the leading edge. Second, they're not nearly as well tested as mainstream hardware. On a platform with thousands of installations you're much less likely to run into a problem nobody has seen before than you are on a platform with only dozens of installations.
How does DARPA know that the money granted unto these respective companies will be used for their intended purposes? Surely some method exists to make sure the money (our tax dollars really) doesn't wind up being another form of government subsidy? (a.k.a: corporate welfare?)
-Pizaz
Was the reason they went away, can't see things have changed when, especially when you can now cluster much cheaper alternatives
Why do people buy those really expensive supercomputers, when they could just buy an Apple one instead? They're much cheaper!
Other posters have already pointed out the bandwidth issues over and over, so I'll skip that obvious difference.
The fact is that not all problems are suitable to parallel processing. Sometimes you really need to know the outcome of one operation before you can go on to the next.
Beowulf clusters really suck on problems where that applies. Cray style supercomputers shine on them.
=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Friends don't let friends enable ecmascript.
Don't just think about solving a static problem faster, it's also about solving a problem better through the use of more variables. Take weather simulation. If having too many variables stretches todays forcast into next week, then it's useless. So you limit the amount of variables to come up with a "close enough" forcast in a more timely manner. With a faster computer, you can get a more accurate simulation in a more reasonable time period. This increase in accuracy/complexity is then useful in many fields.
What is a supercomputer? Look here for an explination. What can we use them for? Curing evil diseases.
:)
These are available here. This might give you ideas as to what kind of staff works on one of these machines.
I'd also like to point out that Cray has been vigorously searching for qualified individuals. I wouldn't know this officially; I check their site out every month or so. Positions get filled quickly!
Oh hell, just read their site. It has enough information about what they do
www.sitetronics.com/wordpress
...isn't 'Cray' today about as 'Cray' as the company that now owns 'Atari'? What's left besides the name of the original company?
Does it hurt to hear them lying? Was this the only world you had?
Power *usage* perhaps......
Have you ever actually run a supercomputer?
You know, that's kinda funny, since it's my current job. ;) I'm a NERSC employee. :P
You're right, until the the system hits maturity. Our T3E before being retired had a lot less hardware problems than our linux cluster does. Or the SP3 we have for that matter.
BTW, since it's rather hard to find a job these days for some people in the computing realm, we're hiring.
Do you know why the road less traveled by is littered with the bones of the unwary?
The name Cray is synonymous with speed and high end performance. Of course many clued up folks are building their own solutions with the power of clustering.
OK this is about as much a kiddy thing as how many VWs fit inside a football stadium or something, but... ...anyone know of a site with info on how current and past supercomputers compare to current desktops? Where are we at now with 2GHz G5s and 3.3GHz P4s, relatively?
One of the comparisons made when I was at university was of a 30-something MHz 386, with a supercomputer from 1973, showing how they do about the same amount of processing/data transfer but in completely different ways. I found that fascinating
Well, almost. Let's say I have a plane that can accomodate 100 people and does NY->London in 6 hours.
My problem is that I have to move 1000 people from NY to London
Now I can either:
1. I can buy a plane that is 20 time faster, 20 times more expensive. That's the supercomputer
2. I can buy 9 other planes (same as mine) and accomodate the same results as in 1 for less than half the price (I'll let you do the math). That's the cluster.
3. I can buy a plane that has a capacity of 1000 people. That's the parallel supercomputer. But if that one can do the deal for my specific problem, it proves to be not that flexible if my problem changes (ie: 500 people NY->London and 500 people from NY->LA).
That's the power of the bewolf cluster!!!
Write boring code, not shiny code!
The oldest Supercomputer in the top100 is the good old asci blue mountain.
Its at place 30. 5 years old now.
Asci red (before the p2overdrive upgreade) would still be place 36. 6 Years old.
And those 2 are only present because they were the fastest of their kind.
For the record: The fastest machine 10 years( the cm5 from thinking machines) has only 20% of the performance of the LAST entry of the current top500.
15 years ago there was no top500 list, but at that time a cray ymp or hitachi s820 were the best of the best. Performance compareabloe to a dual p4xeon or dual opteron workstation.
btw: simulations arent some crazy thing only nuke-builders do.
Some kickass machines of the top500 are doing useful simulation work for companies like basf, gm, bayer, ect.
HI O WISE PRINCE. WHT TOOK U SO DAM LONG?
Probably not. Cray made some money back when a supercomputer was something that an ordinary company might need. The capabilities of "normal" computers was much more limited then today, so there was a much higher percentage of the buying public likely to want something more. These days the vast majority of users are happy with something mainstream
But, you ask, isn't there a lunatic fringe who wants more power at any price? Well, the lunatic fringe ain't what it used to be. During the heyday of cray you got a damn fine box and nothing else. Cray didn't want to worry about your software--or even an OS. A person who needed the speed would plunk down the money for the box and then pay a couple of guys to code everything from scratch. Those days are gone--software is the driving factor these days, and people are far less willing to buy something that's going to force a total code rewrite. Especially if that thing is only going to buy them a couple of years of edge before they need to recode for the next best thing.
Then there's the question of whether cray can afford to be bigger. The answer is "probably not". If you sell to a lot of customers you need a huge support infrastructure. Cray doesn't have much of one anymore, so they'd need to buy one. (Most of the old support guys left one way or another when SGI came in, or stayed with SGI.) If you have a lot of customers you can spread the costs around, but in the case of a company like cray a support infrastructure means having a people sitting around most of the time in every region you sell a machine. Maybe two to four guys per system (24x7, right?) plus some sorta warehouse facility if you enter a new geographical market. That's expensive. You can bill a lot of that cost back to the customers, but that just makes your systems less competetive.
I think the long term answer is that cray will be a very small niche player, selling to a very select group of (U.S.) government agencies, with the occasional pro forma business customer thrown in so the company can issue press releases. Even most government facilities aren't in a position to buy a cray anymore. (Research money is fairly tight, recoding costs are prohibative, MTBF's are more of an issue then they used to be, etc.)
You're simply not going to get that in a single PC cabinet.
In 10 years I might! =P
You can't haul the A-Team around in a Pinto.
Does it hurt to hear them lying? Was this the only world you had?
Oh!.... that Cray!
Never mind!Our T3E was having problems well past the point where it was getting long in the tooth. Cray started adding functionality to make it more supportable a few years back, but when it was actually a cutting edge system it was pretty unstable. They probably couldn't widely sell a system today that had the problems of the earlier T3E's (one hardware problem and you need to reboot the whole thing) but that just increases the development costs and time to market in a market where delay means that the peasents will be nipping at your heels. Remember, by the time a super hits maturity, it's obsolete.
85 replies, even the trolls, and not one "Imagine a beowulf cluster of these" post.
Can't a guy count on slashdot for anything anymore?
--
PCs rule. Why would anyone want to use anything else? This is so beyond my experience I have no comprehension of real computers. I can't even fathom the need for such a machine.
Shutting down free speech with violence isn't fighting fascism. It IS fascism!
My next couch should be a Cray..
Vector processors are much faster, I know, but are they so faster to justify the increased price? Wouldn't it be better to improve scalar processors, or reduce the price of vector processors (like, finding other uses that would justify mass production and consequent reduce the production costs)?
If I was an american citizen, I wouldn't like to see public money invested in vector processors, and would favor the improvement of scalar processors and apps. Leave the market of vectors to the japanese and their playstations.
In SOVIET RUSSIA, Super Computer buys DARPA!
Oh and imagine a beowulf cluster....
- "Scientia non habet inimicum nisp ignorantem"
one can hope they can make a real comeback, but I am not entirely hopeful. Cray has a notoriously cantankerous culture (much of the acrimony when they fell under SGI was of their own doing, as they even refused to get SGI business cards for ages, spending more time being defiant to SGI than anything else - like finding customers). I am curious as to how they are under Tera's custody.
Especially because it's so much easier to hide a computer than an airplane. No sightings in area 51....
We have to assume that the state of the art is way past the public data. Cray has a "lousy" $150 MM in yearly revenue. They could be spending 10X that on heavy computing for national security. The government is spending $25BB on intelligence and another $400 BB on defense every year. Cray could be a drop in the bucket, even a red herring. I'd love to know what is going on in the basements at Fort Meade.
"All that is required for evil to triumph is for good men to do nothing." - Edmund Burke
In the 1970's and 1980's, Cray and other supercomputer companies fit in the niche of "fastest computing at any cost". The design cycles were long for the specialized hardware that pushed the boundaries of the available technology. Companies and government agencies were willing to pay the high price since there was enough processing speed difference between the supercomputers and the "vanilla" computers.
By the early 1990's, the "attack of the killer microprocessors" came. The PC class processors were still weak, but the higher dollar RISC processors used in workstations, like Sun, were reaching performance levels close to what the supercomputers were able to deliver. Since they were based on higher volume and more standardized processors, the price/performance of the RISC workstations started eating into the mainframe and supercomputer market. Many of the supercomputer companies died off, and some started to incorporate RISC processors into their designs. By the mid 1990's I believe that Tera and Cray were the last remaining old-school supercomputer companies left. The rest either died or were absorbed into other companies.
Today, the investment required to produce the fastest processor chips is so high that it requires large unit volumes to pay for the cost of development and production. The PC class processors, with their high volumes, are putting pressure on the old style workstation market, where each company makes their own processor (SPARC/Sun, PA-RISC/HP, Alpha/DEC). We see Sun struggling as the PC's eat their market. Even some large scale supercomputers are based on the PC processors. The majority of the computer spectrum from low to high end is based on the same families of processors (Intel, AMD, PowerPC).
So that brings us to Cray/Tera. Cray seems to go against the economics of scale that drive the rest of the computing industry. What keeps them running is a small niche that the government is willing to keep funded. It is similar to the funding of exotic bombers and fighter jets. We probably won't see Cray grow much larger than they currently are. They be kept running since they form a critical part of the national security, at least that is what the government believes.
God, I could just go on forever. As a compuational chemist, I basically spend my life being clever because the machines at the supercomputer institute aren't fast enough to finish a calculation before I die, or don't have 6000GB of memory to hold a matrix of all of my possible configurations.
But seriously, Cray is about 75 miles south from me. It would be really cool to take a tour of their plant and see the X-1 in person. At best my dual PIII 1GHz machine is good for about 640 MegaFlops, but just one of the X-1 node modules is 50 GigaFlops =) X-1 video
"I bow to no man" - Riddick
Supercomputing per se died because Intel, DEC, IBM/Motorola had a lot more money to throw at speeding things up than the supercomputing community.
In the 70's up until the early 90's it was possible to build a custom CPU out of discrete logic that ran significantly faster than the available microprocessors. Cray was able to push their clock cycle down into the nanosecond range through clever design. However, a 1ns clock rate == 1GHz. You can go buy that multi-million dollar CPU for a couple of hundred bucks in today's market.
In order for superocmputing to be viable you have to be able to provide quantum leap performance above the commodity hardware AND keep your cost/performance ratio in line as well.
The CRAY-1 came out with a clock speed of about 80 MHz and vector processing and high memory bandwidth at a time when mainstream systems like the PDP 11/70 were running at about 7MHz with a 1MB/s memory bus. Microprocessors weren't even't a joke compared with the Cray.
The new Japanese NEC supercomputer came with a price tag of about $160 million if I remember correctly (some estimates say that it took $1G in research funding) and hits 35 TFlops (sustained). #3 on the Top 500 supercomputers list is a Beowulf cluster with 2304 processors coming in at 7.6 TFlops (sustained). Even figuring $2000/processor + interconnect, that puts the Beowulf cluster at around $5 million or 1/32 of the cost for 1/5th of the performance (roughly speaking).
There are other factors, of course, but the key is that for the supercomputer to stay ahead of the microprocessor a boatload of funding is needed for the supercomputer and the payoff just isn't really there. If it was a lot more supercomputer companies would still be in business.
time to market in a market where delay means that the peasents will be nipping at your heels
Bloody peasants!
Actually, I hear ya. The T3E did have some horrible hardware problems in the beginning. In the end, it was vastly more stable. We could run for a long, long time w/o problems. However, the SP3 we have has problems even now. IDK if IBM will ever get the bugs ironed out with this and related architectures...:S Just IMNSHO. ;)
Do you know why the road less traveled by is littered with the bones of the unwary?
"Well, a well engineered supercomputer has much less overhead than a cluster. One superfast processor doesn't have to deal with interprocessor communications like a cluster does."
I like the way Cray put it:
"If you were plowing a field, which would you rather use? Two strong oxen or 1024 chickens?"
- Seymour Cray (1925-1996), father of supercomputing
And how about a few more Cray quotes?
"#3 pencils and quadrille pads."
- Seymoure Cray (1925-1996) when asked what CAD tools he used to design the Cray I supercomputer; he also recommended using the back side of the pages so that the lines were not so dominant.
"I just bought a Mac to help me design the next Cray."
- Seymoure Cray (1925-1996) when was informed that Apple Inc. had recently bought a Cray supercomputer to help them design the next Mac.
I wonder what he's using now? a Palmpilot?
Don't just think about solving a static problem faster, it's also about solving a problem better through the use of more variables.
You're right and all, but the pedant in me can't help but point out that it's not necessarily sucha qualitative difference as you suggest. First, pick the numebr of variables that you want to use. Now, it's a static problem, and the only difference between two machines is how fast each one will solve it.
On th other hand, you've got a good point, in that the difference can be "all-or-nothing" for a given problem if it crosses an externally-imposed threshold, such that the results are only timely for so long, and are only useful if they come within that period. A weather "prediction" program that doesn't complete until after the weather has happened is still intellectually interesting (worth doing for evaluating the post-dictive power of your model, and in anticipation of someday having a faster computer), but you could have just looked out the window.
Oh, and the other way that a more powerful computer can be qualitatively superior is if your dataset / model / algorithm requires more memory than the smaller computer has, so it can't run at all.
David Gould
main(i){putchar(340056100>>(i-1)*5&31|!!(i<6)<< 6)&&main(++i);}
The Earth Simulator proved the power of custom vector processors (over conventional microprocessors). The design of these vector processors are not limited unless backwards compatibility is needed (I don't think that is a real requirement). This means that there are many possible architectures that could be used to implement the vector processor. The largest problem that I have building ASICs (Application Specific Integrated Circuit) is proving the power/area/performance trade-offs between different architechures. It usually takes a couple of weeks to build an architecture to benchmark... it is difficult to keep a team working on target, while listing ideas for the next itteration. No offense against Cray, but Japanese engineers are very good at the itterative design approach (I learned this while doing my MSEE in Japan). Also, I would think that the same engineers who designed the Earth Simulator 3 years ago have been busy in the mean time... learning and improving on the design of their successful processor. Cray has some product offerings, however I would think that they would be limited in their practical design knowledge due to their lack of a prototype on the scale of the Earth Simulator. People can invest in Cray all they want.. good luck! But, even if Ford stock went through the roof, I'm still going to drive a Toyota.
The Sandia National Labs supercomputer (code name: Red Storm), currently being built by Cray, is going to be powered by 10,000 Opteron processors. A 40 Teraflop theoretical peak will put it at the top of the supercomputer list, being approximately 4 Teraflops faster than the NEC Earth Simulator, the current champ.
Palaces, barricades, threats, meet promises
I really want to see cray come out with more waterfall computers. I thought that was the greatest thing in the world when I saw it on Beyond2000! way back in the day. The contemporary "elegant mac" isn't even in the same aesthetic/functional dimension as that cray machine.
Ah, glory days.
Sounds like Cray marketing articles. For example, Daniel Katz at JPL wrote in 1997:
which is > 35% of peak. Or consider this from the Universiry of Liverpool:For sustained/peak of about 60%.
I have no doubt that one could find problems where a Beowulf cluster has 10% efficiency, but there are real many problems that are good to go on a cluster. And even if you only got 10% it would be worth it if the cluster cost 5% of what a vector computer costs. Not to mention that performance/$ on commodity hardware increases by a factor of 2 every 12-24 months. It takes years to develop a supercomputer, and they are stuck at their level of technology for several years since they are so expensive to redesign.
No electrons were harmed creating this post, though some may have been subjected to electrical and/or magnetic fields.
That's like Ferrari selling a cheap-ass subcompact shopping cart. Cute idea for folks who can't get the real deal, but it aint gonna happen. And if it does, don't expect home market pricing - eg, SGI's Intel Workstation affair.
Forget thrust, drag, lift and weight. Airplanes fly because of money.
They bought part of CRAY, the one that made the CS6400 server, which was a really neat SMP system based on supersparcs.
The rest of the company went to SGI.
So basically the server/sparc division went to SUN and then they got the technology for their Enterprise systems.
The rest of the supercomputer (the Alpha based and the Vector based units) units went to SGI, which did.... nothing with them. Oh, yeah they named some interconnections as CRAYlink or something, but they had 0 CRAY technology on them, they just wanted the name.
Same with TERA, they wanted the name and a way of ditching their crappy TM technology.
Moving people in planes is not a good analogy because it is perfectly parallel. Each person getting to the destination is not in any way dependant on the other people's journey, so splitting up the work has no overhead.
The Cray design philosophy is for solving problems that can't be split up easily. If all of the parts of the problem depend heavily on one another, you pay a large price for communication when you split it up. That's the situation where the cluster doesn't do as well as the Cray. So each design has its strengths, and it really depends on the problem.
A processor fails, on average, say once every 10 years.
So you'd expect the CPU in your computer to fry every 10 years or so, if you kept it that long.
The reason you have more issues with multi-processor supercomputers is that.. gasp.. you have MORE PROCESSORS.
Put 1,000 processors in a machine, and intead of 1 failure every 10 years, you get one failure every 3.65 days. And that's just CPUs.
paintball
Do you mean MTBF for a single box, or the whole cluster? One of the big advantages of using a cluster is that you are robust to any particular node failing. So even if your commodity hardware is sketchy, you still have a reliable system on the whole.
Well, almost. Let's say I have a plane that can accomodate 100 people and does NY->London in 6 hours.
:)
My problem is that I have to move 1000 people from NY to London
But what kind of plane would you need if all 1,000 of those people were conjoined, and you couldn't surgically separate them?
I'd bet a plane that could move 1,000 people at a time would start to look a bit more interesting at that point. BTW, I'm leaving the seating arrangement as an exercise for the reader...
[Polka music]
Help-uh. Somebody-uh please-uh, turnoff-uh da bubble-machine-uh!
[Polka music fades away]
My other car is a 1984 Nark Avenger.
Lets say you want to move 1,000,000 pounds of
sand. a fleet of dump trucks might work. But
now let's say you have just on 1,000,000 rock
you need to move. It doesn not matter how many
dump trucks you have it ain't gonna work with
out one really big crawler transporter.
Same with babys. Nine women can't make one in
a month.
Now back to computers: If a prolem is decomposable
you can use a cluster but if you have one big
problem and don't know how to decompose it into
a set of little problems you need to big computer.
Some problems are like babys and big rocks.
but most aren't that's way they seem to build
only about a dozen super computers a year.
Number of TFLOPS isn't everything. The move back to vector style processors in super computing has been largely inspired by the fact that beowulf clusters work really well for some problems - and very, very poorly for others. If you've got a problem that divides nicely into discrete chunks that don't require a lot of interprocessor communication, then yeah, sure go with beowulf. But complex simulation problems have a tendancy to leave most of the processors idling while the cluster talks to itself due to network speed issues.
Why?
Every solution has to be chosen corresponding to any specific need. My point was just to show that in most cases the cluster makes sense. Of course some special cases might be better suited by option 1 or 3.
;-)
you couldn't surgically separate them
How do you stuff them in the plane then?
A good constraint for option 1 would be that you need to have them ASAP and the overall transfer could be interrupted anytime (before the 6th hour) and at that at that time you still want as much people as possible. Let's say 3 hours. Option 1 will have brought half the people there while option 2 leave all the planes above iceland at hour 3 with noone in England.
Write boring code, not shiny code!
SRC Computers is his legacy, not Cray Computer Corp.
He co-founded this company (with several other
ex-Cray employees) and died while still an employee/owner.
Interestingly, SRC is still around without any evidence on their website
of shipping a product. My guess is that their customers and/or investors
prefer to stay out of the limelight.
What a poor choice of acronym. How confusing.
http://yetanotherpoliticalrant.blogspot.com
Disclaimer: I love SGI computers.
Well, from my experience the memory model is what matters the most. In a super computer you can allocate memory the way you need it with less restrictions. Let me explain, I have a data set that is several gigs, in a cluster I have to do lots of magic to ge it to work in a super computer I just say load mydata.dat.
Your trade off here is: Pay a scientist to develop a complicated distributed algorithm, or pay SGI to give you a shared memory machine whichever is cheaper should win, so for fairly parallel problems with small data sets clusters are probably better. Problems with unpredicatble data patterns and large data sets, you have to pay big bucks to SGI.
Secon comes the reliability. Examples: In a 256 nodes cluster at any given time there are 2 or 3 nodes that are not working. My home directory is not mounted in node 34 and then my program fails, I loose one night to run my code. Node 49 which is supposed to have 2 processors only detects one, my code doesn't do dynamic load balancing because that would add another 4 month to the development time, therefore my code runs for twice as long just because one processor was loose.
In an SGI, I get e-mails like the kernel paniced twice, we will have to shut the machine down and apply a patch. Next Monday between 12:00 and 12:30 the machine will be down. Remember, you can checkpint your job and it will start righ where it left or you can just not schedule anything for that time. Even if you are dumb and schedule stuff, we have told it that we will be shutting it down and it is not going to start running anything unless it knows i can finnish. I don't ever recall our machine failing even though it has a 94% usage.
And last, the interconnects are so fast that you can get much more out of each processor. The compilers are very well optimized and all the libraries are there for you to use, already compiled with the proper optimization and ready to run.
Side note: why would anyone choose Opteron for heavy numerical code, there isn't a good compiler for it. I can't understand. AMD are you listening? I need to buy a cluster and I would like to choose your chip but no compiler no purchase.
I've been using Desktop Cray for a while now. It took me some time to weak the settings to perfection, but now it's just running along. Check it out!
/Styx
Well, it doesn't have to be. We could say that a company wants to send 250 people to London and want to use the 6 hours flight to have a corporate meeting in the plane... You're kind of screwed with 10 planes containing 100 people...
In this case option 3 makes sense.
You could say that the 6 hours is a reasonnable limit but sometimes (not predictable) you need as many people as you can in England before (amound of time not predictable either). In this case, option 1 make sense because both options 2 and 3 doesn't deliver anything before the 6 hour delay.
Write boring code, not shiny code!
Yes, but....did you notice that the NEC machine is 5120 processors? It's massively parallel too.
The key, though, is development $$. And development $$ requires either big sales $$ or big $$ from government.
Cray, in 1994, probably one of their better years, had revenues of $920 million and spent $140 million on R&D. Cray was always the best funded of the supercomputer companies, however when you're trying to develop fast silicon $140 million does not go far.
Considering the fact that your chart only shows AFAIK the change in value. It doesn't really tell us much. But I can play this game too:
SCO vs Cray vs DJIA
By that Chart, SCO must be 1000x better then the Dow Jones!
Your hair look like poop, Bob! - Wanker.
I'm no expert on High Performance Computing. But I know a few, and none of them think that Vector Computing was killed off by government policy. Maybe they're right, maybe they're wrong. But John Markoff presenting just one side of the argument as Proven Fact is pretty pathetic. Some tech journalists don't seem to know anything except what they were told at their last Dog and Pony Show.
Tom Ridge won't care but Tom Clancy might be very interested. Sounds like another asinine use of the DMCA is on the way...
Laws are for people with no friends.
Traditional supercomputers, with their vector processing, are more suited for situations where the results of each element in the data set or experiment do depend on the value or outcome of one or more other nodes.
To explain this, it helps to assume that the average experiment is analyzing what happens in a certain field, and that this field can be divided up into discrete cells. Further, I'll assume that for any experiment looking at how something changes over time, you need to divide up the expeiment's timeframe into a series of windows, and you need to know how each cell in your field is being transformed at each stage in the timeframe.
So for example, if you wanted to model what happens in a nuclear weapon or the core of a star, the results of every cell & every timeslice are highly dependant: you've got material moving rapidly across the field, disrupting what would be happening across myriad pathways, and pulling out any one cell for independent analysis is probably meaningless because the environment in that cell is so chaotic. Problems like this are really only suitable for traditional supercomputers.
On the other hand, consider the analysis of weather patterns. You can calculate each point in your field more or less independently, but before moving on to the next timeslice, each node has to exchange data with all its neighbors. A situation like this probably wouldn't be a strong candidate for Beowulf computing unless you can come up with a clever approach, or the network bandwidth of sharing all that neighbor data will kill your performance.
Then you've got problems like, say, analysis of markets or elections. Here, it's probably safe to assume that most actors in the market act independently, and the amount of data exchange among elements is low enough not to be a severe bottleneck.
The trick is to analyze carefully the nature of the problem to be studied. The more the elements in the problem are interdependent, the less appropriate the Beowulf approach. On the other hand, the more you can use Beowulf, the cheaper the experiment, which is probably why the approach is used more than it is perhaps suited do be doing. As in everything else in life, the economic angle has to be part of the decision, even if what's cheapest is sometimes the opposite of what's most appropriate.
DO NOT LEAVE IT IS NOT REAL
Too bad he wasn't as knowledgeable about cars, or he wouldn't have bought the SUV that killed him.
Did you know you can fertilize your lawn with used motor oil?
It gets down to what kind of problem you are trying to solve and how you are trying to solve it. Clusters built from common of the shelf components need an interconnect for dividing up the workload. For algorithms that don't require a lot of message passing between nodes, a relatively low bandwidth, relatively high latency interconnect like Gigabit Ethernet or (compared to something NUMA) Myrinet is sufficient. There comes a point, however, where the amount of message passing and/or the speed of the interconnect may flatten out performance. That is, simply adding more nodes to the problem yields little or no performance gain. You are kind of stuck when this happens, unless you replace the processors in the cluster with faster ones or someone comes along with a faster connector.
A supercomputer (and to a lesser extent large SMP systems, like the p690, E15k or SuperDome) is designed from the ground up with an extremely high bandwidth, low latency interconnect. This means that those problems requiring message passing between processors will execute a lot faster than the self made cluster (in theory anyway). In fact, you might find that such a system with lower speed CPU's will outperform a cluster of 3Ghz PC's, just because the machine is designed with "total performance" in mind. There are other smarts built into the system and the O/S is generally engineered specifically for the processor (e.g. Unicos), which is the opposite approach of Linux and *BSDs.
Which approach is right for solving *your* problem gets down to how much money you have, the algorithm you are using, the size of the problem set, your own expertise and a bunch of other factors. It's a bit like deciding whether to use a semi-trailer or a group of Subaru WRX's to move a house full of furniture from one place to another. Either way will work, but one is generally more optimal for a given problem set.
For a more thorough exploration of this topic, I would recommend the O'Reilly book: High Performance Computing by Severance and Dowd, which discusses these kinds of issues in great detail. It's a bit old these days, but much of the information is as valid now as it was then.
Most of the posters against supercomputers argue relentlessly based on "economies of scale". Aren't we geeks? Isn't there aesthetic value in building a really specialized machine simply to smoke through a small set of problems? Why the blind love for more and more Intel or AMD processors?
I mean, if sameness were the thing that mattered, then, why not just give up the whole Linux crusade and accept Windows and all Microsoft standards for everything? I'm sure we could get distributed processing software working on Windows, for scientific applications. Windows has sockets, it has languages, one could build a networking computing application on top of it...
Let's remember that it's not just necessarily about solutions, its about acceptance of alternative approaches and a desire to unmask the unknown.
This is my sig.
John Markoff, the same jerkoff that wrote the less then factual articles and book about kevin mitnick, and happens to belong to one of the less reputable media outles (aka the plagarized and false stories coming from the ny times).
Lawyers, MBA's, RIAA? A jedi fears not these things!
Wall St. can't buy it, whatever it is... Cray Inc has more shares outstanding than Cray Research did in its heyday approaching $1B/yr sales. Anybody on Wall St. who thinks this stock is going up like the old Cray simply hasn't done their homework. As other posters have pointed out, vectors are cool and have a place but way too much of the everyday supercomputer work can be handled by clusters and such. They have a niche and its cool but don't expect it to grow like the last Cray did.
The really frightening thing about Cray is the people in control (Seattle) built a computer that doesn't work (Tera) and the people not in control (Mpls / Chippewa Falls) are generating all the revenue with their boxes that do work. Too bad they have to carry Burton Smith around on their backs.
No insider info here. You can find all this and more in the annual reports. Happy reading.
However it happens, it is unlikely Cray was wrong about Gallium Arsenide -- he was not stupid. The question is when will a bureaucratic organization be able to throw marching morons at the problem and make it happen -- since that appears to be the only way technology is funded anymore.
It's unfortunate Seymour allowed Cray, Inc. to keep his name after he left to found CCC. Even though Cray himself was capitulating to massively parallel silicon in his final days -- he did die almost immediately thereafter.
PS: It seems creepy he died in a "jeeping accident" -- because that's exactly the way I had portrayed him dying in an April fools joke faxed to all members of congress a few years before -- an "accident" following shortly on the heels of CCC being taken over by Craig Fields of DARPA. I was sending out the joke because of the horrifying way DARPA had spent money on silly favorites within the academic community while guys who were really pushing the envelope like Seymour were going begging for customers -- having acquired private investments.
Seastead this.
The CPUs the beowulf clusters are all vector
processors too. SSE2 and all that.
-I like my women like I like my tea: green-
In heaven, they all have Zauruses.
-I like my women like I like my tea: green-
> td ots processors are often idle because the memory subsystem cannot feed the processor fast enough.
This is an argument for building a faster memory,
not a faster CPU.
> Some map best to vector machines.
Fortunately, all the major high-speed COTS
processors are vector machines. (SSE2, Altivec...)
> The architecture ought to match the problem set.
So write all your code in VHDL and run it on
FPGAs.
> Commodity hardware goes kaputt
I'll take the quality control that goes into
millions of chips from thousands of wafers
over two drunk guys with a soldering iron
and an atomic microscope any day.
-I like my women like I like my tea: green-
Most big problems have big data. But to exemplify
the kind of problem your describing, consider
computing the digits of pi.... You need to compute
the initial, lower precision digits before you
can refine the answer, thus adding digits.
But wait, as it turns out there is an algorithm for
computing any digit of pi independently of the others.
I think that there is no problem that is not
soluble by a massively parallel method. After all,
we are trying to compute results about the real
world -- and reality itself is ultra-parallel.
-I like my women like I like my tea: green-
I think there is one single reason that the market is poised for a Cray comeback... HEAT!
Commodity PCs managed to push the speed envelope by pushing the heat envelope... That's the main reason AMD took the speed advantage, because they were willing to operate their processors at higher temperatures than Intel would at the time.
Now, I would say it's quite a different story. First off, processors are getting closer and closer to the end of the line for heat increases.. Pretty soon, no known metal will be able to conduct heat away fast enough to allow computers to operate at room-temperatures. Even now, dumb little personal computers need serious cooling solutions... Either that, or they need to be some place that has serious air conditioning.
So, what are companies going to do, even with the current line of processors? Should they invest loads of money in dispersing waste heat, powerful air conditioners, system cooling fans, and software and/or hardware to closely monitor temperatures? OR Should they invest in a higher-end system that doesn't put off so much heat, doesn't use up so much electricity, etc?
In fact, I think we are even nearing the point where home users are going to get seriously pissed off and start demanding lower-power systems... It's interesting that C3 processors have become so popular despite their lowsy perfomance... (Maybe AMD/Intel will learn something from that)
So, I do think that either commodity processors will hit the heat ceiling, and stagnate like the rotational speeds of current IDE hard drives, OR the electrical and major cooling requirements of commodity processors will become too much to justify the small price savings. Either way, that will leave the market wide open for serious computing companies once again. The only question really is how much longer will it be until one of those two things happens? Well, in the Southern California Desert, electricty prices are still very high, and the temperatures are so very high that running a modern computer 24 hours a day requires your home cooling to also be running 24 hours a day, just to operate within the heat tolerances. I don't think it will be much longer before more of the country, and the world, will reach the same point.
Slashdot gets worse every day... Pipedot: News for nerds, without the corporate slant
I guess you have never heard of things like Sun's SunFire range - CPU failure in your 106-CPU 15k? No problem. Just map out the faulty board, remove the board, add the new board, and map it back in all whilst the machine is still up and running.
Same with all other components - hotswap memory, CPU, PCI cards etc etc etc.
It just works.
People should not be afraid of their governments - Governments should be afraid of their people.
Lot of people playing down vector processors here today without having read up on what they are, don't lump vector-processors in with general purpose mainframes and such.
Sounds like nerd heaven, usability hell.
There is no afterlife, so stop reading Slashdot and get on with it!
That was classic intercourse!
It's not just more variables, although that is a factor, it's about how many gridpoints you can use to represent the physical domain your model covers. For instance, today's predictive models which cover the day-to-day weather over North America or Europe use too coarse a grid to "see" thunderstorms - they just don't exist because the spacing of the gridpoints is too large for such a "small" phenomenon to occur within the model. Faster computers with more memory will allow forecast models to be able to see the atmosphere more realistically from small scale features to large scale features.
As a meteorologist with lots of modeling experience, I can say with a lot of confidence that you simply can't throw too much computational power at atmospheric modeling. Personally, I won't be happy until our global models are running at 1 centimeter resolution up to the thermosphere.
A squid eating dough in a polyethylene bag is fast and bulbous, got me?
It will be banned by the government as you'll be able to mod chip it to control cruise missiles...
-- Leeeter than leet
For a nuclear winter.
retrorocket.o not found, launch anyway?
I tend to agree...
In past lives I've worked with career spooks and even had a few of my own visits to the NSA (no I've never been to Fort Meade - any real spook groupie knows that the Infosec guys in "X" were up at the "Friendship Annex" by BWI...).
But it was an infrequent but not common practice at vendors that some systems to certain customers would be delivered by leaving a trailer at location X where it would disappear for a few days then they would get a call to pick up the trailer at Y.
There is also a verification process to sell to private and foreign customers to verify that systems aren't going to the wrong people.
But the real question is that now that Cray is Cray again will they start shipping a case of Leine's with every system like they used to? (Probably not given that manufacturing is no longer in Chippewa Falls, WI).
Come now, they may be short vectors, but they are
every bit vectors. And yes, I've written CAL code.
And vectorizing compilers.
-I like my women like I like my tea: green-
One has to draw the line somewhere, sure they are vectors but a 32 bit word is a vector of bits as far as a binary and instruction is concerned, no one could imagine calling a binary and a vector instruction though.
Strictly speaking I admit that your position probably is the much more common one when it comes to the terminology though :)
So, what is the real win on vector operations? When I was first learning about vector architectures in the late 80's, at first I thought that the advantage the vector processor brought was that it crunched the whole vector simultaneously. I was very disappointed when I discovered that the majority of implementations had a max of 4 arithmetic units.
:-). What I learned is that the win for vectors (unless modern units have a boatload more processing units) is that the memory I/O is much faster into and out of the registers, there is no loop overhead and accessing the registers by the arithmetic units is very fast.
As part of my learning experience I was tasked with writing a simulator for our next generation processor so we could start fudging benchmarks early
When I was at FPS we were discussing an i860 based array co-processor. I wonder if anyone has thought of building a vector unit around a set of Pentium or PowerPC chips? They could emulate vector instructions with the registers being held in a shared cache. With proper coding there would be no need for interlocking and the "registers" (emulated by the cache) should be able to blow data in and out of an interleaved memory system as fast as any regular vector unit would.
As for the "it's worth it if it only costs 5% of the vector machine", for small clusters that may be true. For large clusters, they actually cost very close to what our vector machines cost. All those Myrinet/Quadrix switches aren't cheap either...
Go Badgers! -- #include "std/disclaimer.h"
I hate to break it to you, but Intel and IBM/Apple/Motorola are, in fact, using vectors to increase the speed of their commodity processors. Just because it's not in the Intel x86 instruction set doesn't mean it's a bad idea.
Go Badgers! -- #include "std/disclaimer.h"
No electrons were harmed creating this post, though some may have been subjected to electrical and/or magnetic fields.
It's not caused by anything.
Back in the olden days, Ctrl-H was also the command code for backspace.
^ means control.
Basically it's the net-speak version of "VA Linu.. er.. Software. *Wink Wink*"