Slashdot Mirror
SKA Telescope To Provide a Billion PCs Worth of Processing
Sharky2009 writes "IBM is researching an exaflop machine with the processing power of about one billion PCs. The machine will be used to help process the Exabyte of data per day expected to flow off the Square Kilometre Array (SKA) telescope project. The company is also researching solid state storage technology called 'racetrack memory' which is much faster and denser than flash and may hold the secret to storing the data from the SKA. The story also says that the SKA is unlikely to use grid computing or a cloud-based approach to processing the telescope data due to challenge in transferring so much data (about one thousand million 1Gb memory sticks each day)."
A thousand million is probably the most correct term for international understanding.
There is no world standard term for one thousand million. In the US and most of the UK we call it a "billion", but in several countries a billion means a million million. In these countries, a thousand million is usually called "a thousand million" or a "milliard", but I've never seen "a million thousand".
It's be roughly 105,882,352,941,176,470,588 discs.
.5 in thick but encased for storage probably 12*1 which would be about 12 tapes per cubic foot
They are about 10.5 in in diameter and
The Volvo V70 has about 72 cubic feet of free space
About 122,549,019,607,843,137 Volvo V70 Station Wagons... or one making the trip 122,549,019,607,843,137 times *shrug*
They charge processor per socket not per core.
Maximum diameter of 10.5". Tape width was 1/2". The protective ring around the tape reel would add about another 1/4" to the diameter, and the thickness of the reel sides and retaining ring would add about another 1/4" to the thickness.
We're not going for a rigorous space-filling solution; we'll stack the tapes in a square array (rather than, say, a hexagonal one). So the tapes effectively become 10.75"^2 x .75" rectangular prisms. That's 0.0501573351 cubic feet per tape.
According to this scan of the 1972 Mercury station wagon brochure, the 1972 Montego MX had 91.6 cu ft of cargo space. That's 1826 tapes. (Assuming dimensions of tapes and station wagon are compatible and don't leave some wasted modulus.) So 1,000 tapes is in the right order of magnitude, but a smidge low. Good guess. But your "number of cars" division was off by a factor of ten. Using the correct version of your numbers, 58,820,000 tapes transported 1000 tapes at a time is 58,820 loads. Using the 1,826 tapes per load number, it works out to 32,213 loads.
I calculate that the capacity of a station wagon full of 9-track tapes works out to 310 GB. (170 MB per late-era IBM 3400-series tape reel at maximum length, 32K blocking, and 6250BPI. At least that's what Wikipedia says.)
Hmm... I wonder if the various IP performance calculations would work out for a MTU of 310GB and a ping time of minutes to hours (depending on trip length and freeway speeds)....
Welcome to the Panopticon. Used to be a prison, now it's your home.
Some people might be interested in knowing where all this data comes from. There's a rule of thumb in astronomy that the angular resolution of your images is the wavelength of the radiation you're receive divided by the diameter of your telescope. Radio wavelengths are pretty long (up to tens of meters), so you need really big telescopes, which you get by scattering lots of little telescopes all over the place and then looking at the how the phase of the incoming radiation shifts based on location. So what you do is you sample the voltage of each antenna with 1 or 2 bits of resolution at the Nyquist frequency. So for 100Mhz radio waves you sample at 200MHz. That's 50MB/s for a single antenna. The SKA will likely have tens of thousands of little antennas scattered all over the place. So say 50MB/s times 20,000 antennas = 1TB/s = 100 petabytes/day, which is about what the summary says.
Now, it's not quite as bad as it looks. You don't have to pipe all this data to a central point to analyze it. You can take a small group of antennas and just look at the correlations between those, combine the data from that group and send the combined data to a second level of correlators, which takes data from a set of small groups, and so on, in a hierarchical fashion. You lose some information this way, but you get most of it, and the only wa to get all the information out of the data would be to bring it all to a central processing location so that data from all antennas could be compared to that of all other antennas, which is O(N^2) in the number of antennas and obviously infeasible for a telescope like the SKA. Even as it is, the system of hierarchical data collection is really pushing Moore's law, as the article shows.
average english football field of 110 x 67.5 meters (74,250,000 sq cm)
1,000,000,000 memory sticks of 4cm x 1.5cm each (6,000,000,000 sq cm)
About 80 football fields.