Slashdot Mirror
Jaguar, World's Most Powerful Supercomputer
Protoclown writes "The National Center for Computational Sciences (NCCS), located at Oak Ridge National Labs (ORNL) in Tennessee, has upgraded the Jaguar supercomputer to 1.64-petaflops for use by scientists and engineers working in areas such as climate modeling, renewable energy, materials science, fusion and combustion. The current upgrade is the result of an addition of 200 cabinets of the Cray XT5 to the existing 84 cabinets of the XT4 Jaguar system. Jaguar is now the world's most powerful supercomputer available for open scientific research."
There is a queueing system. If you want to run a job on a machine like this, you log into the control node (which is just a linux box) and submit your job to the queue, including how many CPU's you need for it and how much time you need on them.
A scheduling algorithm then determines when the various jobs waiting in the queue get to run, and sends mail to their owners when they start and stop.
On many machines there is a debug queue with low limits for number of CPU's and runtime, and thus fast turnover; this is used to run little jobs to ensure everything is working right before you submit the big job to the main queue.
Each project has an al
Assume every single person on earth can do a 16-digit operation on a calculator in one second.
It would take them roughly a quarter million years with no breaks of any kind to do what this machine can do in one second.
Ultimately human behavior is near-continuous series of yes/no decisions. Our brains iterate pretty deeply, but at some level it's ones and zeros. Though we may need more petaflops than angels on the head of a pin before we can scratch that itch. At any rate, the application of such a model will probably always doom it to failure.
How much do we really know about climate? Probably a lot less than we think. Scientists are always so sure they are right. And then a few decades pass and they realize they weren't. And then they repeat that same behavior.
Help fight poverty: Punch a poor person.
You, sir, are an idiot.
LANL, LLNL, and SNL are all weapons labs. ORNL is primarily a science lab.
I myself have worked at three of these labs and held an account on an earlier iteration of Jaguar as well as some of LANL's other supercomputing clusters, so I ought to know.
ORNL's Jaguar cluster, although parts of it are I think "controlled" rather than open so that it can run export-controlled code, is not at all classified. It's used for biology, astronomy, physics, CFD, etc.
Also, if you knew the first thing about classified security you would realize that disallowing FTP access on a *classified* (Red network) machine to the outside internet is a necessity. To my knowledge, they don't allow *any* interconnection between classified systems and unclassified.
I had but a simple dream, to destroy all humans.
remember that its not a true 64-bit multimedia system its two 32-bit systems connected together XD
Having done my graduate work in fluid dynamics, only half your statement is possibly correct. There is historical evidence for global climate change, both warming and cooling. If is our interest to maintain the current status quo, the climate as we know it, it is not at all clear what interventions we need to take, or what effect they might have. Without simulations, that correctly model the real world. We have no way of knowing what our interventions might do. If anyone is interested, I can elaborate. The short (and scary)answer is - resonance.
Every time you have sex outside marriage, god kills an economist (and a kitten)
NCCS is a capability site, so no. You just need to be willing to wait for your job to bubble up to the top of the queue. In fact, as a capability site, the whole point is to develop codes that can run on the entire machine. Now, once your job runs, you will have to wait a while to get another opportunity, as the queues are set up to provide an 8 week moving average of "fairness".
Check out the gallery if you haven't.
I've always wanted to get some custom graphics like that on my server racks. Maybe a penguin, a butterfly, and a can of Raid. :)
Supercomputers definitely don't look as exciting as they did in the "old days".
No. Jaguar. 1995 XK12, Six-Litre.
I'd rather doubt the 1995 XK12, as cool as it was, was any competitor to the Jaguar XJ220.
Damn_registrars has no butt-hole. Damn_registrars has no use for a butt-hole.
Los Alamos are jealous since they just got a 1.026 petaflop supercomputer installed earlier this year by IBM called RoadRunner. It was featured in last month's Linux Journal.
this nation, under God, shall have a new birth of freedom. -- Lincoln, Gettysburg Address
I believe Cray made its partnership with AMD quite a while ago while they were still ahead of Intel in the performance/power ratio. In addition, these machines have a very fast interconnect (SeaStar) that is based on HyperTransport links. I believe it was recently announced that Cray has formed a partnership with Intel, and I imagine they will port the technology to QuickPath for future machines, but QPI was not available at the time this machine was commissioned. One does not simply order a machine like this at the drop of a hat. Vendor decisions are typically made years ahead of time.
creating better algorithms? Or at least educating a little bit all non-CS scientists about performance and optimization?
The guys who work on the the dynamic cores for the biggest climate models (NCAR, GFDL, NASA, etc.) do world class numerical hydrodynamics. Maybe not quite on par with the nuke guys at, say, Sandia, but pretty good. And they do hire programmers and numerical methods people to do algorithm design, optimization, and parallelization. They're cutting edge in terms of grid solver algorithms for these sorts of problems. There are lots of complications from irregular topography, coupling between atmosphere, ocean, biosphere, and cryosphere, etc.
If you go over the climate-prediction loop many many times, you should consider some caching..
Caching does little good, because none of the grid cells have the same value after each time step. Some things just require a supercomputer. There's a reason why people use supercomputers for big 3D fluid dynamics simulations (nuclear explosions, virtual wind tunnels for aerospace, climate/weather models, etc.)