Slashdot Mirror
IBM Constructs New Fastest Computer
scoobysnack writes "MSNBC is reporting that IBM has once again created the world's fastest computer -- it will be used for simulating real-world nuclear tests. With 12 teraflops it would still take it 3 months to simulate the first 1/100th of a second of a nuclear bomb explosion." There's coverage at CNET as well.
So, simple math being your friend, let's assume that we're talking June, July, and August. That's 92 days. Each day has 86,400 seconds. So that gives a total of 7,948,800 seconds in June, July, and August. Now, this amount of time only simulates the first .01 second. So we need to multiply by 100, yielding 794,880,000. This means that the computer needs to be 794,880,000 times more powerful to simulate the first .01 second of a nuclear explosion in .01 second.
Multiplying this figure by 12 (They said it had a peak of 12.3, but it probably can't do 12.3 all the time), a computer would need to be capable of 9,538,560,000 teraflops to model the first .01 second of a nuclear explosion in .01 second. And the complexity increases from there.
Damn.
Would you rather they did open-air explosions to do their tests?
You mean a packet with the 'OOB' (out of band) flag set... right?
How come almost every time there is a post about supersomputers, they are being used for nuclear bomb explosion simulations?
As usual, there's a good article over at The Bulletin of the Atomic Scientists on why there's so much government desire for bomb simulation.
We have treaties and treaties on why we can't test these devices "for real". Given the desire to upgrade them without "upgrading" them in a way that affects counts or treaties, there's currently a lot of interest in how to re-use existing designs and components in ways that give functionally new weapons, without being listed as such. Converting air-burst devices to near-surface burst devices turns town-killers into bunker-killers, but it doesn't have to appear as building new weapons or changing the type of existing ones.
By the way, the weather service did get themselves their own parallel computing cluster (running Linux, by the way). Incidentially, the progress made in simulating nuclear blasts carries directly over to astronomers who simulate supernovae.
a Beowolf cluster of these?
Seriously though. This is using the Power III-3. Isn't the INSANELY fast Power IV just around the corner? When will that sucker arrive?
Actually, they could hand them out to the public.. these are not tactical simulations, but actual particle simulations.
I believe the problem arises in the amount of shared data required between nodes... it's not like cracking a key where you can just chop they keyspace up into as many pieces as you like and work on them all separately.. you have to have the entire dataset in order to work on it properly..
They aren't working to improve the explosion. It works well enough already.
They are ensuring that the bombs will go off reliably. The thinking is that if we have a working nuclear stockpile, enemies will think twice before attacking us. If our weapons get old and fail to work, the deterrence will be lost. How do we know if our old weapons will work? Blow one up for real, or simulate it.
I'd rather them simulate it.
If tits were wings it'd be flying around.
Should this article be given (-1: Redundant)?
They already used 'Blue'. Twice.
Asci Blue Pacific (Livermore) 1999
Asci Blue Mountain (Los Alamos) 1998
And Red.
Asci Red (Sandia) 1999
So, being American.. it's time for white, yes?
I went into the machine room for a major animation company during an interview and among the racks of Origin 2K's and disk arrays they had
a couple of midrange Cray's.
I assumed they were for graphics processing (that's what the Origins were for), but it turns out they were simply the fastest fileservers on the market at the time of purchase (an important thing if you're pushing around mutli GB files).
--
"L'IT c'est moi!"
efficiency.
More kill for the buck.
Screw that where's my damn flying cars???!!!
There aren't nuclear weapons simulations per se because such things require tuned parameters from actual tests and are not directly useful for anything other than making nuclear explosion, but there are open source (mostly public domain actually since they were government sponsored) particle simulations which can be used for things like simulating the propogation of radiation in the human body.
--
"L'IT c'est moi!"
IBM Constructs New Fastest Computer
IBM's ASCII White Super Computer Unleashed
It's CmdrTaco coming down the stretch on New Fastest Computer, but here comes timothy on ASCII White, it's Taco, it's timothy, Taco, timothy....timothy by a nose!
--
Linux MAPI Server!
http://www.openone.com/software/MailOne/
(Exchange Migration HOWTO coming soon)
Apparently, you need to release a similar amount of energy to simulate in real time...
to get an idea of the scale of this, the whole SETI@Home project is generating about 8 TeraFLOPS. This thing tops that by about 50%. So it could process about 500000 SETI units per day. or just under six *per second*.
:)
Keywords: Quake 3, Kernel compilation, Beowulf, Toy Story 3 in realtime?
Fross
The supercomputer industry is fading. Only the government is willing to shell out big bucks for top end machines. They only have a handful of applications that qualify- weather prediction, bomb simulation, airplane design.
Large scale distributed computing over heterogenous networks are only suitable for computation that can be broken into discrete computational work pakets.
This is exactly what Distributed.net and Seti are doing, as far as sending a discrete block of data, having the client crunch it, and then return the result.
Some computations require a network of interdependencies int he data set. Such as large scale simulations, which is what this computer will be doing. In such scenarios, there is no way to 'break up' the computational tasks into neat little discrete packets, since they are interdependant.
This requires lot of very fast networking (ccNUMA, very large SMP, etc) on hardware designed specifcally for this type of task.
When you get down to it, a "new" (just off the assembly line) nuclear device is relatively controlled and predictable. It won't blow up until you tell it to, and you know how it'll react when you push the button.
As they age, though, you don't really know what'll happen. And that's why we have simulators. As others have pointed out, there's only 2 ways to know if a device that's been stockpiled for 15 years will work - simulate it or take it out to the desert. Now, since we can't exactly take one out back and set it off, we buy a bigass computer and simulate it.
Or would you prefer that one just randomly go off while sitting at the dock inside an Ohio-class submarine at Groton, CT? Or maybe, if we return to the 50s mindset of 24-hour alert for bomber crews fully loaded for WWIII, with the occasional scrambling to test readiness, that B2 flying over Topeka hits some turbulance and levels half of the already-flat state of Kansas?
Actually, in a sense, this is a "beowolf cluster" - it's a big collection of separate machines with high-speed interconnections. It probably doesn't use "Beowolf" technology, however. Read the press release.
...phil
...phil
"For a list of the ways which technology has failed to improve our quality of life, press 3."
You know, more goes on in a nuclear weapon than fission. If we can properly simulate the beginnings of FUSION, that could be an important step towards commercially viable fusion power plants! Cheap, clean, unlimited energy... worthy goal, I would think.
Additionally, even if the data from this box *IS* indeed ONLY ever applied towards nuclear weapons, that's still MUCH better than the alternative: which is to withdrawl from the Test-Ban Treaty, and start setting the things off for real again.
SIMULATING something is NOT morally equivelent to DOING that very thing. Otherwise, quite a lot of Quake, Carmagaddeon, and GTA players would be sitting in jail right now.
And, hell, even nuclear bombs, as they exist now, and as they could be refined, have potentially non-military uses. No, I'm NOT talking about Teller's harbor in Alaska; or the ridiculous scenarios in Deep Impact or Armagaddeon... Although nukes COULD be used for the noble purpose of deflecting incoming Comets/Asteroids. The implimentation, as presented, just sucked.
Actually, what I'm talking about *WAS* mentioned in Deep Impact. I'm talking about the Orion drive. If we are ever smart enough to withdraw from the ridiculous treaties which prevent it's deployment, Orion could be the answer to all of our short term space exploration problems! Until we perfect fusion, it IS the most powerful drive system proposed for deployment. Imagine how FAST we could get to Mars, and how much equipment we could take along if we used Orion, rathar than ridiculously inefficent chemical rockets!
Or, for the peaceniks out there... Wouldn't that be the ULTIMATE "swords into plowshares" situation? Imagine... the nuclear stockpiles of the world, ultimately directed not towards mutual annihilation, but towards the exploration of the final frontier!
We HAVE the way, all we need is the will.
john
Resistance is NOT futile!!!
Haiku:
I am not a drone.
Remove the collective if
Imagine all the people...
Isn't the idea of these aggregate computers (clusters, whatever) that you can just keep on growing them?
Add a couple hundred nodes, buy another switch or two, increase your flops by a couple of hundred g?
Johan
These things must be the latest fashion in international peeing contests - It used to be that the US was upset that the USSR had enough missles to blow up the US 20 times over, and we USians could only blow up the USSR 15 times so we (USians) had to make and deploy more missles to acheive 'parity' and get the USSRians back to the negotiating table.
Now-a-days, I guess the US is afraid that China will have better nuke simulators than the US so we gotta beat 'em at it, it's "Keeping up with the Chin's" all over again.
I'd rather see the funds go toward a modern super-collider but, pfft, I only pay 1/3 of my income to taxes, I don't have any real say in how it's going to be spent.
try { do() || do_not(); } catch (JediException err) { yoda(err); }
The nuclear simulations do help in the stockpile stewardship program, by stopping those large creators. Also, from what two lab officials working heavily on the NIF project (and worked on past projects) told me, much of the work allows scientists to keep the stockpile updated. Old bombs become dangerous, and the government ruitinely signs treaties requiring that new techniques must be created. The advancement in the research alone helps numerous industries, and building the machines of course fuels that technological research. So, its not all that horrible, but this research (in treaties) help stop other countries from conducting nuclear tests. Simulating them in the lab is far better than on bikini island.
"Open Source?" - Press any key to continue
Nuclear simulation, like fluid dynamics, is basically a cellular simulation -- make several bazillion cells, time step each one, communicating only with the neighbors on each step.
(I think. I'm actually making this up as I go along, so add salt to taste)
Now the problem is that since the entire simulation goes in lock-step, you limit the number of steps by not only computation speed, but also communication speed between the nodes.
I presume that there are smart approximative approaches that can be used to assauge this, but it remains the case that distribution and cellular simulation just don't go.
Johan
Isn't the idea of these aggregate computers (clusters, whatever) that you can just keep on growing them?
Add a couple hundred nodes, buy another switch or two, increase your flops by a couple of hundred g?
Any idea if that's what they did? None of the articles have said whether they just plugged 9 more teraflops worth of power into the existing 3 or put a whole new 12 tf system in. That would be interesting to know.
Kintanon
Check out JoshJitsu.info for Brazilian Ji
Gamingmuseum.com: Give your 3D accelerator a rest.
Most people who ramble on about NP have no clue what the N actually stands for. Your question has no real meaning in the sense that it asks about an attribute that is not associated with the question (rather like saying "How many doors are on the dog?"). The simulation is polynomial w.r.t the number of particles being simulated, but exponential w.r.t to the mesh granularity; it's entirely practical to approxiamate systems of non-linear differential equations to as much precision as one wants to wait for. In this case, a more powerful computer means that one only has to wait months instead of years to get results starting from physical first principles (instead of experimentally derived hueristics).
NO!!! The operating principles are as removed from a beowulf cluster as a bicycle from a car("Yes, they both have wheels, gears, and are made of metal."). The RS/6000s are not commercial off-the-shelf parts; they are substantially modified to allow greater internode communication bandwidth and lower backplane latency. The network switch for SMP (known as an interconnect) is orders of magnitude faster than even gigabit Ethernet; it is entirely custom built and accounts for a majority of the developement cost of the machine.
The OP's comment is utterly moronic. Solving ODEs and PDEs is a task perfectly well suited for a conventional computer. Quantum computers do *not* double their performance each time a single atom is added; not even when a single qubit is added to the device. Obviously, you are a troll or otherwise attempting to exploit the moderation system.
Yeah the Japanese machines are probably TCM based sysplex units a-la older IBM type ES9000 mainframes. I believe the largest off the shelf TCM machine is a 12-way. These processors are enormously fast, consume tremendous amounts of electricity and throw off vast quantities of heat which is why they're water cooled. If we compare the performance of TCM units vs. the latest CMOS mainframe class CPU's it still takes about 3 CMOS to match the raw performance of 1 TCM. Now moving down the scale, the IBM-like mainframe class CMOS CPU's themselves are built specifically for mainframe machines and have very very high performance baselines. How high? Hard to tell since IBM will not publish performance benchmarks for mainframe machines that can be compared to other types or brands. Instead they use an internally derived benchmark that uses a 'commonly' know basic performance figure based for example on some well known IBM class mainframe like a 9021-831 or something like that. Any other machine is evaluated as a factor or that. At any rate the latest mainframe class CMOS machines have complexes or the rough analog of SMP cages that contain at least one CPU (up to..I don't remember, you can check). Each complex or base machine model is then sysplex'd to other same-type machines up to 12 or 14 machines or even higher. This is what the Hitachi/Fujitsu machines do. They build an x-way complex and then sysplex all the complexes together. As a rough comparison a VERY large commercial sysplex is typically a 12-way with each complex containing 12-24 individual processors for a total of 144 to 288 discrete CPU chips. This honestly is the high end of the high end for standard (non custom built special purpose) mainframe class machines. Compare this to an IBM RS/6000 SP2 frame with say 8 nodes of 12 processors each and ganging 20 or 30 or more frames together across a second level backplane switch for a total assembly of at least a few thousand discrete CPU's to more or less the same work. At least in the commercial world. In the nuclear simulation world obviously you want the highest possible FP performance so a TCM based mainframe design is enhanced with additional or different vector processors compared to simply exploiting the general purpose FP performance of whatever RISC CPU you're using. Another reason why the numbers of CPU's in the two classes of machines is so different.
Here's a link to an article, but's it's a bit dated:
http://www.ibm.com/news/1999/12/06.phtml
=Blue(23)
LITTLE GIRL: But which cookie will you eat FIRST? C. MONSTER: Me think you have misconception of cookie-eating process.
Dear Citizen,
We have built this giant computer to simulate Nuclear Explosions. Previously, we couldn't predict the outcome of a Nuclear Explosion. We did not know if it would kill a few million people, or a few billion. Until we had the ability to simulate it we couldn't be sure, and if we aren't sure, then we can't protect you. So please continue to send us more tax dollars to support the electric bill for our new Nuclear Explosion Simulator(TM) and we can continue to protect you. Also, it's good for children.
On an unrelated note, please feel free to update your PGP keys to the longest possible key length you can use, we believe you have every right to your privacy.
Yours Truly,
Big Brother
-----
On a more serious note, how much ass would we kick if we could get this badboy to join Team Slashdot over at distributed.net?
-Tommy
"I got a half gallon of Jack, and 2 dozen Ant Traps. I'm about to get wild." -me
If they had posted an article (or two) stating that the U.S. was going to resume above-ground testing, how would that make you feel?
This is the "alternative" that they are always talking about in those debates. So, quit whining or we're gonna have to make Nevada glow.
#VRML V2.0 utf8
For all intensive purposes, "whom" is no longer a word. That begs the question, "who cares"?
Since the other story is gone from the home screen, jump to it and steal all the +5 stories and repost them here for free karma!
How come almost every time there is a post about supersomputers, they are being used for nuclear bomb explosion simulations? While I realize that this is a better thing to simulate than to actually do, aren't these computers being used for anything else? Is it that the people who these computers are being built for only want them for those purposes? I just think it would be great to see an announcement mention that a supercomputer would be used for analyzing weather patterns, help with the human genome mapping effort, or something else, well, different. :-)
Ok, What I want to know is where did the old computer go? They had a 3.??? teraflop computer before. Now they havea 12.??? teraflop computer. What did they do with the 3? Scrap it? Give it to another branch of science? Sell it? Stick it in a warehouse? Why can't we take it and set it up in a big room and let every research facility around that wants time on it buy some. Or even just allocate X amount of time per month for each scientific institution and let them use it to further research. It would very much suck if they just threw the thing away....
Kintanon
Check out JoshJitsu.info for Brazilian Ji
Well, it's an IBM computer and I don't think IBM would really want to be using chips designed by SGI/MIPS. SGI has constructed ASCI Blue Mountain (a 3 teraflop machine) and has submitted a bid to build a 30 teraflop machine for the next phase of the ASCI program.
No. The interactions between particles during the initial phases of a nuclear explosion are highly nonlinear and often not in local equilibrium. The interactions are orders of magnitude more complex than the gravitation force; there is no known way to recursively block and calculate aggregate forces and effects. Besides, the objective to simulate nuclear explosions from first principles means that the usual simplifying hueristics cannot be used; one cannot substitute simplier equations (if an explicit form even exists, which they usually don't) for the systems of differential equations that need to be solved.
You need internode bandwidth and low communications latency; this type of simulation can only be done with a large, monolithic memory space machine.
"How the power of this computer compares to that of Distributed.net or similar projects?":
A comparison is not really possible. d.net can do more raw operations per second if one just adds up the total power of all the machines involved. However, ASCI White has far better internode communication. For the purposes of during highly paralizable calculations like distributed FFTs on blocked data and brute force sieving or key searches, d.net is probably faster. For the purposes of doing tasks like numerical linear algebra, nuclear/non-linear dynamics, weather forcasting, etc. ASCI White would be faster.
"How feasible is the distribution of such a computation? Are all the calculations similar, or would a lot of different computational code have to be written?":
It would essentially be impossible to distribute the computational task, even if the intial value data could be distributed (and it can't; even simplier FEA simulations routinely have datasets exceeding 30 GB), the process would require so much internode communication that any d.net type system operating over a heterogenous, low-bandwidth, high-latency, public network like the Internet would never work.
"Are there any such systems already in place? Currently, I'm only aware of one "useful" system, and that's ProcessTree (damn, I lost my referral number). SETI@Home is arguably useful, depending on whether you believe there is extraterrestrial life that uses the same radio waves we're scanning and is sending signals we could interpret.":
Not for using distributed computing for these kind of tasks.
IBM's ASCI
Draws 1 Point 2 Megawatts
The West Coast Goes Dim
"It's tough to be bilingual when you get hit in the head."
And my vote for worst processor name in current production: IBM's Power3-III!
Jeez, get some imagination ya nerds.
Hotnutz.com - Funny
And how come they only give us the flops count?
how about more information like frame rate?
__________________________
"Oh, you hate your job? There's a support group for that, it's called everyone, they meet at the bar."
I believe it is time for us to reconsider the role of the military. After WWI (or was it WWII) we renamed the War Department into the Defense Department. This reflected a more peaceful mindset. Since we haven't had any defense of land and life in the US for over 50 years, it seems appropriate to rename it to the Foreign Economic Interests Department, since the military is used as a pawn to secure American economic interests.
Of course, this is a separate matter again, but certainly there is something better to simulate than an explosion. What possibly could they seek to understand about it other than how to improve it? This being the last thing society needs.
-- Solaris Central - http://w
OK, you've already been moderated down (which is
good), but honestly--what else would you have them
do? Play Quake at 50k frames/sec.? Crack crypto
keys?
'Nuclear simulations' doesn't just mean figuring
out how many people they can kill per megaton.
It ties in plasma physics and a dozen other
related fields, which tie in closely to
astrophysics, i.e. how stars happen. Then there's
the residual interesting bits, which can lead to
advances in almost any random field. (maybe not
random, but indeterminate)
Fundamentally, they're using tons of computing
power to investigate stuff we _don't_know_ yet.
They're not just cranking out bigger numbers
faster, but looking in different directions.
To say, "I don't think we need more (fill in the
blank) research" is to utterly fail to understand
how good science works, and ties together.
"People who do stupid things with hazardous materials often die." -- Jim Davidson on alt.folklore.urban
the difference between the 16 way nodes in the ASCII white and a standard RS/6000 node is largely a matter of packaging. they are half-width 4u rackmount boxes instead of standalone or full width. you can buy the same (electrically) 16way smp node as a standalone webserver or workstation.
the switch is what puts the super in this supercomputer. but when you break it down, it is just a fast network. yes, orders of magnitude better in all ways than 10base-t but still just a fast network.
and as far as operating principles, it is just IBM's flavor of MPI. nothing special there. all the money went into the switch.
maybe the horsepower equation deserves the bicycle/car comparison, but the operating principle is the same. i.e: a bunch of standalone unix nodes, connected by a high speed network clustering software and MPI (or PVM).
The difference between Theory and Practice is greater in Practice than in Theory.
With these computers, you can test nukes without having to actually blow them up.
A deep unwavering belief is a sure sign you're missing something...
This is a nice piece of kit to say the least, who wouldn't want one for themselves, but look at the use it's being put to - running simulations of nuclear bombs being used. Yes, it's again part of the $1 trillion USian milatary machine, even if it is given a more publicly acceptable face through the Department of Energy's "Stockpile Stewardship and Management Program", a euphamism if I've ever heard one.
Why is it that the Pentagon still gets to spend so much money on fancy new toys for a war that will now never come? The USSR has collapsed and since the US is now sucking up to China it looks like there isn't going to be the proposed World War III that US military leaders have been hoping and planning for for decades. And whilst this $1 trillion goes into the military black box, poor people starve on the streets and can't afford even basic health care thanks to the Randite social policies of the US, despite what their Constitution supposedly guarantees.
No, on purely technical merits this computer is interesting, but I don't think that we, as reponsible people, should be praising something which is part of a group that contributes in a large way to the suffering of the poor and needy.
---
Jon E. Erikson
Jon Erikson, IT guru
Blue Mountain (at LANL), for instance, was on the order of 6,000 RS10K processors; if you hunt around enough, you can still find the webpages about it at the Lab. (Again, I'm lazy. Sorry. :-)
Not only that, but the POWER chips are WAY WAY faster than the latest MIPS.
A deep unwavering belief is a sure sign you're missing something...
Sorry to rant a little. My point is just this: argue all you want that we shouldn't have nukes; write your congressmen, campaign on Capitol Hill, etc. I wish you luck. But until that day comes, it makes sense for us to do this sort of simulation.
This machine can run 100 hours in a row? It takes 2 hours to boot? It takes an army of white suite programmers to support? Damn! I was hoping I could use it to check my email, but now...
You can't handle the truth.
Of course, depending on the number of procs, you get different results with the different -j options
make -j2 zImage
make -j4 zImage
make -j zImage
etc...
"It's tough to be bilingual when you get hit in the head."
The day you realized that atoms, too, had subparticles, that was an epiphany.
The day you realized that splitting an atom would release megatonage of energy, that was an epiphany.
The day you realize that it would take over 25 years to simulate one second of the blast in a computer, that was an epiphany.
It's a new kind of physics, you need a new kind of software.
"boot up quake 3"? You scare me, my friend. I'm sure some people would like to run quake as their OS, but...
Gotta agree with that. I also want to know what they are hoping to discover about the first 1/100th of a second anyway. We know what happens, there's a bright flash, shit loads of heat, and lots of people die, either immediately or later. What else is there to know about a nuclear explosion. Surely if they're going to spend this amount of money on a supercomputer, they could put it to better use. Bill Gates could always use it for his bubble sort, eh? ;-)
Now weary traveller, rest your head. For just like me, you're utterly dead.
At least we can run our own weather simulations at home with the Casino-21 project. How long until a distributed nuclear simulation project? I guess that wouldn't happen becuase of "security concerns," though.
Prevent email address forgery. Publish SPF records for y
No, a 3D game is just about the most stressing thing you can do to the computer.
A) A GeForce2 GTS can render a hell of a lot more triangles than the proc (even a 1GHz) can feed it. Sure the geometry acceleration helps out a bit, but most games don't use it yet. Thus this racing game (no racing games that I know of use the geomtery engine in D3D or OpenGL) is definately a good indicator of the performance.
B) The kernel compile is a crappy benchmark. Given the fact that the source tree is some 75 megs, and the fact that it does nothing with the FPU, and the fact that it is much more dependant on bus bandwidth due to the nature of the operation, it doesn't make for a very good benchmark.
But in the end, all that matters in a benchmark is how well it does what YOU'RE doing. If you want to test raw proc speed, you'll use a synthetic benchmark that just does math ops. If you compile all day, then the kernel compile is a perfectly valid benchmark. If you run 3D games, then the 3D game is a great benchmark.
A deep unwavering belief is a sure sign you're missing something...
Some aerospace critics lay some of the blame of recent rocket failures on just this point, that too much emphasis is being put on rocket simulation at the expense of actually building prototypes and testing them. Certainly it is cheaper to simulate them, but you can't skip too many prototype iterations in the design phase.