Slashdot Mirror
Iran Claims Two New Supercomputers
dcblogs writes "Iran President Mahmoud Ahmadinejad announced development of two supercomputers Wednesday. Iranian government news media published a photo spread of one the systems it claims is capable of 89 teraflops, which is far short of the petascale systems in the US and China. There's no independent verification of Iran's claim. But after the Stuxnet attack, Iran may be trying for an IT comeback via supercomputing or just trying to show it is in control as regional unrest spreads. Iran says the new systems will make the global Top 500 supercomputing listing, but it hasn't submitted a Linpack benchmark to the list organizers."
Iran has a halfway decent computer industry, they even make computer games:
http://www.questofpersia.com/main/index.html
"MIT betrayed all of its basic principles."
Because the FIRST place you think of when you hear "supercomputer" is, of course, Iran.
A simple oversight I'm sure. But then we have a state controlled media vs a non state-controlled Linpack so...
I work for the Department of Redundancy Department.
Ok so I can understand why they would want to have a supercomputer or 2. They must have a lot of computational work being done (building nuclear weapons is pretty technically demanding stuff, so I have heard). But of all the reasons to have a supercomputer the summary states.
But after the Stuxnet attack, Iran may be trying for an IT comeback via supercomputing
Comeback via supercomputing? What does that even mean?
Violence is the last refuge of the incompetent. -- Isaac Asimov
like, how fast can it run StuxNet?
I'm sure it's to process all the different permutations on how to bring about the 12th Imam.
Life is not for the lazy.
Those are SuperMicro servers. I resell supermicro and as much as I love their low cost and good speed, the hardware failure rate is astronomical. They should fit in well with Iran's centrifuges. :-P
The Iranians have, somehow, discovered how to use the global trade in commodity parts to build a cluster computer(consisting of suspiciously under-filled racks, with a bunch of generic 1-3U-looking compute/storage nodes, much more empty space than I would have expected, and some pretty ragged ethernet interconnects, no visible brand IDs; but the black with reddish handles on the drive caddies looks a lot like de-branded HP...)
Other than perhaps minimally-puncturing the (always false) notion that Iran is a bunch of ignorant sand-dwellers just because it is a theocracy, I'm not seeing the big deal here. Depending on the CPU/RAM specs and how many racks there actually are(the photos are fairly cagey on the subject), it very much looks as though they've managed to put a few million dollars worth of datacenter together. News to anybody who thought that Iranians spent their time wallowing in backwardness and squalor; but pretty low-rent by cluster computer standards...
The Steve is not impressed.
Only a few tiny modifications to this otherwise dead on observation.
* Nodes don't need video, and nearly any server/cluster motherboard will have onboard cheap video anyway, so save the video cards.
* "Rack them" is a bit more than Infiniband and parts off of NewEgg. Right idea, but there is a bit of work involved in selecting racks, rackmountable chassis that will fit your motherboard and provide adequate power and cooling, mounting your motherboards by hand in the chassis, and racking them up. Not difficult but not for idiots either, and it is usually a lot easier to buy ready-to-plug-in chassis mounted nodes from the many companies that sell them (with service contracts).
* There are some subtleties in the "Install Linux" part as well. There are "cluster specific" distros that support certain kinds of supercomputing, and more general off-the-cuff Fedora or whatever. Either will work, but if you want to run a centrally managed supercomputer with the ability to do reasonably fine grained computation, there is a bit of work involved in matching up your hardware, network, operating system, cluster environment. Again, not difficult but shall we say "Expert Friendly" -- at the very least you have to be a solidly competent Linux/Unix/Network admin and sytems geek to assemble, install, and manage everything.
So there are a few hard parts even before writing the software. How hard it is to parallelize your software depends on what it is and what you want to do with it. If all you want to do is run a bunch of Monte Carlo simulations in parallel to get good cumulative statistics, it's bone simple and you don't need any fancy version of Linux, OTC Debian or Fedora or whatever will work fine. If you want to solve tightly coupled astrophysics problems involving massively distributed solutions to sets of coupled ODEs, well, that's a real chore and you have to think about it BEFORE you go buy all that hardware at NewEgg.
Then there is the infrastructure -- power and AC -- which might well be a problem in Iraq. It won't do to have the power go down just before your three week long computation finishes, and AC failures can destroy your entire cluster overnight.
The point is that building your own beowulf, while easy enough for geeks, is still a bit of an engineering chore and works better if you have experience and/or expert help. The cluster in the picture looks like it is the size of a midrange University cluster in the US, hardly even worth submitting to the Top500 list.
rgb
Even when the experts all agree, they may well be mistaken. --- Bertrand Russell.
As far as I know, don't we still have a trade embargo with Iran? If so, how are they using Supermicro servers and Windows 7? They are both U.S. companies.