Slashdot Mirror
NASA Upgrades Weather Research Supercomputer
Cowards Anonymous writes "NASA's Center for Computational Sciences is nearly tripling the performance of a supercomputer it uses to simulate Earth's climate and weather, and our planet's relationship with the Sun. NASA is deploying a 67-teraflop machine that takes advantage of IBM's iDataPlex servers, new rack-mount products originally developed to serve heavily trafficked social networking sites."
...what are they doing to improve the algorithms used to calculate the results? And if they're transparent (e.g. open for public inspection) - bonus!
(yes, I know that there are only a few folks in the Human race that would even know how to read the things. That said, it would be nice to have something educational, and at the same time open for public scrutiny so as to avoid political accusation, you know?)
Quo usque tandem abutere, Nimbus, patientia nostra?
Does this mean that the forecasting simulation for tomorrow's weather will run in less than 24 hours?
NASA Engineer: "You know, chief, I've been thinking. I bet we could just about triple the performance of this thing if we supercooled it."
Manager: "Super what?"
Engineer: "Chilled it to absolute zero, like in the large hadron supercollider. Speeds up the electrons."
Manager: "What would you need to do that?"
Engineer: "Oh, I don't know, maybe... a ton of liquid helium?"
Oh, yeah, it's not easy to pad these out to 120 characters.
Seeing as 67 teraflops is going to be the new processing power for this machine, I wonder if a NVIDIA CUDA implementation has been considered. Their Tesla systems are designed for this High Performance Computing, offer a significant amount of processing power and are relatively easy to parallelize code for. I know that oil companies use these high powered systems to find locations of oil, but I guess that its less likely for weather forcasting since there is less money in it. However, it would be interesting to see these cards used for modelling hurricanes and determine their expected strength and path of travel more accurately.
Faster does not mean better. I'd rather have less iterations per day on a good model than many of a crap model.
Engineering is the art of compromise.
I just saw that today, perfect timing
Si vis pacem, para bellum! For evil to succeed good men need only do nothing!
No wonder they're not getting anywhere replacing aging shuttle fleets if they are playing with rubber ducks and earth climate modelling.
Engineering is the art of compromise.
Forecast, not predictions.
Learn the difference. And compared to 30 years ago, they are a lot more accurate.
The Kruger Dunning explains most post on
Even if there is some randomness, we can still try to predict short-term events (this week's forecast)
Yeah. Would you choose a neurosurgeon who pokes around people's brains in his spare time? I wouldn't.
NASA= National Aeronautics and Space Administration
NOAA=National Oceanographic and Atmospheric Administration
Administrations are just that: administrations, run by administrators - primarily to the advantage of amdinistrators. Adding another level of administration won't fix that.
Engineering is the art of compromise.
When an xhpl score says '67 teraflops' and nVidia/AMD gpus spout off about the ludicrous number of gigaflops they have, it simply isn't the same.
For example, the PS3 variant of the Cell processor claims 410 gigaflops. It's hpl score, however, would be about 6-9 gigaflops. Even the new cell processors 'only' get 200 gigaflops by xhpl count.
32-bit precision scores aren't comparable directoly to 64-bit operations.
XML is like violence. If it doesn't solve the problem, use more.
Use rubber duckies in the cooling pool?
Sometimes boldness is in fashion. Sometimes only the brave will be bold.
Ducks (G&R)
You are being MICROattacked, from various angles, in a SOFT manner.
Is tripling the performance really that news-worthy? Either they update their hardware infrequently (every few years or more) and are falling behind everything else in the computer industry, or they are doubling or tripling every year or two, which makes this not an uncommon event. It seems like orders of magnitude would be a scale of real accomplishment - or even better - if they tripled the accuracy of the weather predictions.
While there may be no Blue Screen of Death, it does appear that it is a Big Blue planet.
Even with chaotic systems, and especially with systems where physical laws, especially conservation of energy and momentum, things are pretty predictable for certain time ahead. Chaos comes from uncertainities adding up, until finally that flapping of butterfly can be deciding factor of hurricanes direction. But once a hurricane is moving in a direction, it has a lot of energy and momentum, and it'd take a motherload of butterflies working togheter coordianted to change it's direction... Or one really big one ;-).
And that's weather. Climate is another thing, but pretty much same things apply, except in global scale and over longer times and with different uncertainities.
Forecasting chaotic systems can be summed up in one word: inertia.
I wonder if they take into account the heating that toy generates when making those calculations! ^_^
what, does this one run server 2003 HPC?
Good people go to bed earlier.
Isn't this small potatoes to the power of the distributed project climateprediction.net?
duke out
Not to knock NASA (I'm rather fond of them), but what are they doing in the weather prediction biz, anyway? Last I checked, weather and climate studies were in NOAA's domain.
Cross-agency collaborations are great and appropriate, but in general I'd just as soon see NASA's budget dollars stay invested in space research.
Listen to what I say, not what I mean...
iDataPlex? Really? I am a tester at IBM. We've just started to qualify various hard drives and IO cards for the iDataPlex systems. They're very oddly designed and in general suck. The firmware (BIOS/uEFI) is really crappy but it usually is at this stage of testing. I'm sure it will get better over time. The thing that most likely will not get better is the horrible, horrible physical design (which was specially request by Facebook). I would say the reason is unknown, but from what I've heard it's because Facebook didn't want to upgrade their racks/rails so they had IBM design servers to fit them.
There's lots of curious and pointless design features. They're almost like big-ass blades, designed to slide out of a larger outer-housing that contains the PSU and fans, but several cables and wires connect the machine to the outer-housing making it impossible to remove without also removing the outer-housing from the rack. In one variant, the pci-slot is literally in the middle of the system (imagine a card slot in the middle of your motherboard, that, when a card is inserted into it, acts as a locking bar).
All the ports are in the front of the system: vga, usb, ethernet. Except for power. Power is in the back, attached to the external shell. There are also ps/2 ports (a rarity among newer servers) but they are completely blocked by the faceplate.
My overall reaction: meh.
Skiffy is Spiffy, but Ort is tort.
As much as I hate the thought, your funding is about to go bye bye.
I am very small, utmostly microscopic.
plz provide a fps spec for Crysis. ~:-)