Slashdot Mirror
First Petaflop Supercomputer To Shut Down
An anonymous reader writes "In 2008 Roadrunner was the world's fastest supercomputer. Now that the first system to break the petaflop barrier has lost a step on today's leaders it will be shut down and dismantled. In its five years of operation, the Roadrunner was the 'workhorse' behind the National Nuclear Security Administration's Advanced Simulation and Computing program, providing key computer simulations for the Stockpile Stewardship Program."
RR must've spent too much time pecking on that Acme birdseed.
it's be interesting to see if this thing goes for scrap value, or if someone else'll pick it up for service elsewhere...
Operation Guillotine is in effect.
Oh wait, that's PEBCAP.
"Sound barrier" was and remains OK because there is a physical difference between flying slower than and faster than the speed of sound. But the word "barrier" is now (over)used to make things sound more dramatic. Raising a number from below to above some arbitrary (usually number base-dependent) threshold does not imply crossing a barrier, unless by barrier is meant "barrier to entry of another over-hyped tech piece".
I think that thing could host a kick-ass DOOM session!
---------------------------------------
Rotate the pod, please, HAL....
I bought my aluminum iMac back in 2007 and it works just fine, though I wouldn't mind an upgrade. What happened to that damn sequester?
On another note, can you really use a computer to accurately calculate how fast a nuclear arsenal will deteriorate? I didn't think so.
"In total, Roadrunner takes up 278 refrigerator-size server racks, and connects 6,562 dual-core AMD Opteron and 12,240 Cell chips. "
So way less than Google's Map Reduce computer systems. Why not just buy time on map reduce?
Also these supercomputers are not nuclear stock pile inventory systems (or simulation systems as they put it), and certainly weren't used to help find a cure for aids. They're for SHOW not for use, the stock piles have not changed, yet the computing power needed to track them is 200 times more? BS. And you need to keep re-rerunning the same simulation for a static system? BS.
And home come IBM's *super* computers are benchmarked (well Linpack anyway), yet they use legal contracts to prevent anyone benchmarking their overpriced mainframes?
Instead of spending silly money on a supercomputer for a lab that doesn't use it, why not build a big distributed computing center for universities to use as they wish?
I worked on Roadrunner. It was a pain to program, but I'm sorry to see it go. The Cell processor was ahead of its time...
http://nnsa.energy.gov/ourmission/managingthestockpile
"Most nuclear weapons in the U.S. stockpile were produced anywhere from 30 to 40 years ago, and no new nuclear weapons have been produced since the end of the Cold War."
And yet you need an ever increasing super computer to track and simulate them???
Then we get to the unpleasant truth:
http://nnsa.energy.gov/ourmission/recapitalizingourinfrastructure
"Recapitalizing Our Infrastructure The FY2011 Budget Request increase represents the investment needed to transform a Cold War nuclear weapons complex into a 21st century Nuclear Security Enterprise. "
It's a boondongle, a broken window, a way to turn printed cash from the Federal reserve into a thing that pumps money into the US economy. An agency with the role of spending budget on a vague list of tasks.
The reality is very dull, USA doesn't build Nuclear power stations, the nukes haven't been expanded since the cold war, the 'terrorists-with-nukes" is security theatre, They get a huge budget, they don't have stuff to spend it on, so they buy a supercomputer from a US vendor, which is only IBM these days. They run benchmarks on it, take pride in it, and then do nothing with it because it has no purpose.
They could instead build a really useful large distributed computing system for Universities across the USA to use, that would be a productive use of Fed created magic money.
At today's prices, I'd have it farming Bitcoins.
* Carthago Delenda Est *
you have "flop" in the name
Table-ized A.I.
Posting Anonymously on purpose...
Roadrunner was a failure from the beginning. We learned quite a while ago that there is more to a supercomputer than just hardware... you have to have software that uses that hardware. Buying into exotic architectures without thinking about the consequences that's going to have for the people creating the software that needs a supercomputer is a terrible idea.
Roadrunner was as bad as it gets in this regard... proprietary accelerators that caused you to hardcode your software directly for running on that machine and none other at all.
But, it appears that we haven't learned our lesson: http://hothardware.com/News/Solder-Problems-Trip-Up-Titan-Supercomputer-Delay-Final-Certification/
At least Titan's Nvidia cards can be addressed using OpenCL and/or CUDA... so there is some hope that if they ever get it to work properly they might have some software to run on it...
I guess it's not new and shiny anymore, so we can just throw it away.
I did want to read the actual article, but the only link is to a 2008 article.
Fail or what?
http://news.sky.com/story/1071902/supercomputer-pioneer-roadrunner-to-shut-down
That is the article. And i see why they are getting rid of it, not as power efficient as new computers.
Be seeing you...
Also, they should use a proper car analogy and use "Milestone".
And Matlab does very similar simulations every day on humble PCs:
http://www.youtube.com/watch?v=TvlIfSlLB0c
I myself do a mass of finite element simulations on hardware that isn't even good enough to run old versions of Crysis.
So essentially this thing was a real life WOPR.
"The second simulation (of a full nuclear fuel rod in 3D) was nearly 300 million degrees of freedom and the output alone was nearly 400GB to postprocess. It involves around 15 fully coupled, nonlinear PDEs all being solved simultaneously and fully implicitly (to model multiple years of a complex process you have to be able to take big timesteps) on ~12,000 processors."
And it didn't need to be. I can use 1000 times more nodes in a FE analysis and soak up power too. Why would I? That would be dumb! Your simulating a simple heat transfer and simple expansion, NOTHING MORE, no different that any other chemical process simulation in any other factory. Just with a lot more nodes.
It's also an arbitrary simulation serving no purpose. You said "what is that panel is broken right there' then ran a simulation with a stupid number of nodes to soak up a computer. But the pellet was made, it exists, it didn't need your simulation to be made and the simulation make zippo difference. You can run any number of similar simulations with the damage in an infinite number of places or combination of places, and it makes zip difference to the world because you don't know where each pellet is damaged. So NONE of your simulations apply to the actual pellet.
Their mission statement is absolutely clear. Turn cold war spending into security theatre spending and that's your job.
". The essential part of those is that they're radioactive. That means they decay. So yes, they do change over time.... the supercomputers are used to simulate the decay process and firing to see if they still work "
Indeed, so predictable we make atomic clocks out of them. Of course it doesn't take a supercomputer to make an atomic clock because radioactive decay is simple math.
"Are you somehow under the impression that these supercomputers are used to count nukes and keep track of their addresses?"
Keep in mind you're paraphrasing *their* mission statement in a comic way. That is what *they* say their job is.
The sever got too old so it was no longer good for petafiles.
The Coyote finally won
Yeah, that too, is possible !
Muchas Gracias, Señor Edward Snowden !
http://www.norstad.org/matrix-multiply/
You can reduce matrix multply across many nodes just as you can split it across many cores in a super computer.
Which OS was it using? Linux? AIX? Something else?
"No need to be so rude"
Google *is* your friend. You can paralyze matrix multiply, and indeed you need to, since modern supercomputers are parallel beasts they need to run parallel code. If your algo doesn't scale then your supercomputer doesn't scale.
"the overhead for Hadoop is still much higher than the overhead for cpu's in a supercomputer to talk to each other"
Lets put this in context. The super computers used to design and simulate cold war missiles has less processing power than a modern dumbphone. Even the Cray's of the 80's, as the cold war was ending, were 8core x330mflops = 2.6 Giga flops, or about 1/30th the power of a modern desktop PC.
Each *NODE* on a map reduce network is 30 times is more powerful than the WHOLE supercomputer used in the original design could EVER have been and internally each node has far faster ( x10000) interconnect between the cores. Obviously I'm talking about CPU nodes, if the nodes had a GPU, we'd be talking 1000 times more power still.
GP's simulation was a simple finite element model there, its far simpler than most I've seen. Basically a simple tube, not even pipes, and so on that you normally see in such analysis:
http://www.youtube.com/watch?v=T56-cQ5ZDzc
He doesn't even need to map reduce that, he gains nothing by increasing the number of nodes to a silly number other than to soak up processing power. His data doesn't support that number of nodes and it's the sort of calculation a desktop PC running cheapo Lisa will even do.
And if he wants to run silly numbers of nodes, well no problem because finite element analysis scales nicely and even has Map Reduce implementations. (Google is your friend remember).
Mod point suppression aside, the decay of a fissionable material is not a chaotic systems that needs constant simulation. It's a predictable timely thing, and that's why we can use it for atomic clocks.
Let me ask you, does your simulated warhead work now? Does your simulation say it will work next year? Does your simulation say it will work in 10 years time? Are you claiming that a new computer running the same simulation faster in 10 years will generate a different result?
If you can run the simulation today on a computer of that power, can you run it on a computer of that power in ten years time? Of course you can.
And likewise if you can run that simulation on a supercomputer from the 90's, then it can run on a computer of comparable power in 2013. The problem hasn't gotten 1000 times more complex, it's the SAME problem. So of course you can.
OK, last of the Crays from the 90s was T3e_1350, which could reach 3 Teraflops using 2176 processors, and had 544 GB of memory.
http://www.cs.rit.edu/usr/local/pub/wrc/courses/arch/machines/cray/T3E-1350.pdf
A core i7 will do about 68 GFlops, so about 50 Core i7s boxes, given them 16gb each = 800GB ram.
Or we could also use a Geforce GTX500 and go the GPU CUDA core route, they are about 1.5 teraflops/unit, so a few of those in a single PC is comparable.
Then why are you buying a supercomputer? Why aren't you saving the tax payers money instead? Their mission statement makes the reason clear. It's about spending budget not physics.
They rely on the resonant frequency of atoms in metal vapors (Cesium or Rubidium), or the output of a hydrogen maser.
http://en.wikipedia.org/wiki/Atomic_clock
Radioactive decay is a chaotic process. So chaotic that it can be used as the basis for a random number generator. Just what you DON'T want in a precise time/frequency reference.
http://www.fourmilab.ch/hotbits/
Remember "News for Nerds, Stuff that Matters"? Help make it a reality again! http://soylentnews.org
How do i buy the now "surplus" parts?
Fair point, and if you take a million atoms of a radioactive element, is the half life chaotic? What about a billion? Still random and chaotic? And a trillion? Random and chaotic?
So is your weapons grade nuclear fuel degrading in a chaotic manner or a predictable manner? Of course its predictable manner, the half life is known and unchanged by pressure, temperature etc. you can say its chaotic at the atom level, but you don't have the data to the atom level, and the effect is macro and well defined and stated as half life.
You see the 'like clockwork' rhetorical point is conceded, but actually, the real point stands. It's a predictable system, which is why the can predict it now, just as they could predict it with 1990s supercomputers. Moores law is brutal, the problems stay the same size, but the computers needed to solve them become cheaper and cheaper at an insane rate.
It isn't that the problems magically scale up to the largest supercomputer you can make, and thus you get the budget for the largest supercomputer and get Linflop bragging rights!
Floating-point operations per?
Reading down there is a paper running Linpack on Hadoop and showing it scales excellently.
http://www.st.ewi.tudelft.nl/~iosup/ec2perf-sci-comp09cloudcomp.pdf
And you can see from the pictures that IBM's supercomputer is really no more computing than a datacenter floor of which Amazon, Microsoft (and even Facebook) have many much larger data centers, in their EC2, and Hadoop compatible clouds. They also have far newer kit than the 5 year old processors. Cell processors are not magic, IBM used them because IBM makes them for Sony.
And MS/Amazon are marketing cloud computing.
So wouldn't it be amazing marketing if they cleared enough users off their EC2 cloud, ran Linpack on it and declared their cloud the worlds fastest supercomputer?
I bet Microsoft would love bragging rights over IBM, but even Amazon want to market EC2. It's not a difficult target to beat, there's really nothing special here, its all just parallel linear algebra and Hama has done the legwork.
Also the radioactivity mesh calculation can certainly be done in Mathlab. Wouldn't it be a nice promotion to show the very same simulation done on a humble PC in Mathlab? Calculating the same graphs as his simulation run on the supercomputer.. you won't run 300 million degrees of freedom, but then it doesn't need it, that was just busy work. The output graph will be the same. Sort of "Mathlab, bringing the power of a supercomputer to your PC"... or "PCs are modern day supercomputers", a real marketing win.
a beowulf cluster of these!
I was relatively late to the build your own PC craze, I built my first one in 1995 after about 8 years of being a Mac owner.
But since that time I have found relatively little worthwhile "upgradability" in my systems. I do remember adding a 3DFX card to my Pentium-166 system and replacing a couple of video cards (in the last 2-3 years) whose fans have quit.
When I built systems I tried to get the best bang for my buck out of my CPU, buying just high enough in the product lineup that my parts were "better" but below the point where they wanted $800 for a CPU because it was in the absolute top tier (eg, the P 180s and 200s were much more expensive than the performance gain, and the P133 savings didn't justify the performance loss from P166). Eventually the P200 became cheap enough that I *could* have upgraded, but the cost wasn't worth the nominal performance increase, especially when I was 3-6 months away from a whole new architecture with all kinds of performance improvements (PCI, PII/PIII, etc).
Although that example seems dated, it's always been like that -- there are either limited upgrades (dead end of a chipset/CPU) or upgrades that hardly seem worth the hassle when a new architecture is available with 10x the performance.
I get it that there are guys that chase the latest & greatest video card or who start with the lowest end CPU for a chipset and then serially upgrade via eBay or other bargain hunting, but my interests are more aligned with what's ON the computer rather than upgrading for the sake of a 5% clock rate jump.
Bitcoin server
"No matter where you go, there you are." -- Buckaroo Banzai
Whenever there is a story on supercomputers on /., there will be a comment stating that there was no barrier whatsoever. But that's not quite true.
The truth is that the performance of supercomputers grows that fast because engineers continuously solve problems, which were deemed intractable before (e.g. power consumption, reliability, network performance). The research may not be groundbreaking in the sense of earth-shattering, but definitely in the sense of "wow, I didn't think one could do that!"
Computer simulation made easy -- LibGeoDecomp
Roadrunner consists of 6480 QS22 Blades. Using Cellminer each will yield approx. 56 MHash/s, or 363 GHash/s in total. Using the Bitcoin profitability calculator we can then estimate that one will gain ~27 BTC/day (ATM $2667) while paying $3360/day for power (assuming cheap $0.07/kWh). So yes: mining on Roadrunner would not be cost-effective.
Computer simulation made easy -- LibGeoDecomp
How many FLOPs has it flipped over its career?