Slashdot Mirror
Japan Builds World's Fastest Computer
claylikethemud writes "The New York Times reports that Japan has built the world's most powerful supercomputer from "640 specialized nodes that are in turn composed of 5,104" NEC processors. The machine boasts the computing power equivalent to the 20 fastest American supercomputers combined, and with a top speed of 35.6 teraflops, outpaces the next fastest machine, the ASCI White Pacific, by more than factor of five. Applications include climate modeling, global warming prediction, and other non-weapons research."
Non Weapon research??
Yeah right !
Uh.. from Chapter II, Section 9 of the Japanese constitution:
"Aspiring sincerely to an international peace based on justice and order, the Japanese people forever renounce war as a sovereign right of the nation and the threat or use of force as means of settling international disputes. 2) In order to accomplish the aim of the preceding paragraph, land, sea, and air forces, as well as other war potential, will never be maintained. The right of belligerency of the state will not be recognized."
The Japanese are only able to maintain a defensive force, not an army, so even if it was weapons research, it would only be for use in self defense.
slashdot!=valid HTML
Well, if you look at the number of processor of this supercomputer it's 5104 * 640 = 3.2 Trillion processors.
...) than those kind of computer. Not all applications can benefit of the NEC supercomputer (same things for the Beowulfs).
;-)
I'm not sure that the beowulf approach can follow this performance path. Of course, each processor is more powerfull however, you have to bring electricity and network to all of them (maybe half or a quarter of them if you use multiprocessor motherboard). Here, the number of processor is _huge_ which allow a massive
parallelization of your code.
Of course, some can argue that "we can always build a beowulf of those" but as far as I know, it's not really COTS material and I'm not sure it will be true one day.
Maybe Beowulf will not be able to compete in terms of peak performance. However, for the price it will gives you much more power.
As a conclusion, I would say that Beowulf are less parallel (even with fast network à la myrynet or infiniband or
Today NEC is faster and more powerfull but my guess is that it will not last
The questions are when and how ?
I'm not a real expert but I have recently taken a high performance computing course from somebody who is an expert for my comp sci masters.
The basic problem of adding more and more processors is keeping all the memory in sync. If you have a process that is running across 50 cpus the machine needs to ensure that if one of them updates a variable that all the others work with the current value. (Ok, it's more complicated than that but I'm not writing a book here)
The solution is to write your system so that the calculations can run as independently as possible. However, at 100 million processors it probably just doesn't fit the problem space.
That which does not kill me only makes me whinier
Japanese people are very anti-nuclear-wepons - which is not really a surprise due to the fact that they had two dropped on them. In fact they have sent letters of protest to the heads of every country that tests nuclear wepons since 1965 - hundreds of letters.
The basic problem of adding more and more processors is keeping all the memory in sync.
That's why message passing is typically used instead of some sort of shared memory approach. You eliminate the synchronization problems as well as memory contention. After that, it's just a matter of keeping all the processors busy.
"We returned the General to El Salvador, or maybe Guatemala, it's difficult to tell from 10,000 feet"
Japan has a military budget of 45 Billion. 7 Billion more than France, 3 Billion less than Russia.
The US Budget is 265 Billion.
My problem with what you said, is that you made it seem that Japans budget is close to the US's budget.
All in all, it is a surprisingly large amount for a country that doesn't go into military actions. Who are they defending themselves from?
Democrats or Republicans. They are both taking us to the same place and they are not afraid of us anymore.
Wrong. Just plain wrong. Explicit message passing can often reduce communication overhead compared to coherent shared memory, but the synchronization problems are still very much present. You still can't operate on data before it becomes available, regardless of the programming model. Explicit message-passing systems handle synchronization very differently than shared-memory systems, but those problems don't just go away.
Slashdot - News for Herds. Stuff that Splatters.
Or is it? A Japanese no longer unthinkable
Fortran always used to be the dominant language for vector code but C also (I've been away from this game for a couple of years but did spend the best part of 10 years of my professional life "Vectorising" code.)
The real trick to Vector code is to work with the memory subsystem and not have it permanently trying to catch up with you as in normal processor style.
You can pretty well put as many floating point units in a modern cpu as you want, the problem is feeding them with data to operate on and storing results. Current microprocessors use multilevel caches to try and keep what it hopes are useful subsets of main memory close to the cpu. Trouble is if you are scanning 100's of GBs of data in a weather model there may be virtually no useful small subsets.
For vector processing you design a system where, before you actually fire off any calcs, you give the memory system a list of the next 64/128/4096(varies) addresses you plan to use. It may take a little while to get the first one but after that they arrive at 1 per clock per memory pipe and, depending on the number of memory pipes you use you can actually drive your floating point units full speed.
Because you want to process streams of memory addresses as a single op (vectors) you spend all your time looking for loops where each iteration can be calculated independent of the next and where the compiler can be sure of that with no ambiguity. That tends to mean no subroutine calls, anything a(i)=f(a(i-1)) is bad but a(i)=f(a(i+1)) is fine and even a(i)=f(a(i-65)) can be OK depending on vector register length. You then get into CIGS (compressed index gather scatter) ops like a(i)=b(c(i)) and you can work with that sometimes etc.
Bottom line, if you don't vectorise high 90%'s of your code the Vector computer is a very expensive room heater. You then need to worry about 99% parallel code+ for multinode architecture but there are similarities between data independence of vectorised loop coefficients and parallel modules.
HTH
Crash
Contrary to rumor,
n e02.html
the machine is constructed from 640 nodes, with 8 vector processors per node, and 16GB RAM per node. That totals 5120 processors and 10TB memory.
See http://www.es.jamstec.go.jp/esc/eng/outline/outli
Also of note:
peak performance per processor: 8 GFLOPS
total peak performance: 40 TFLOPS
Remember, when they give you TFLOPS or TOPS values, they're giving you PEAK values.
In reality, most of the time, performance is way below peak values, even for the algorithms for which the computer was designed to handle. IBM's pacific blue has a peak TFLOPS value around 3.6TFLOPS...but in reality, its usually around 1.2TFLOPS.
There's no reason to believe this machine will be any different.
Furthermore, the performance of this machine is likely to sink like a rock when its used outside the area it was specially designed for.
In other words, the best supercomputers in the world are still the ones made by starbridge systems, which were bought by NASA (I believe the one NASA bought was called HAL 15, or something like that).
social sciences can never use experience to verify their statemen
What the heck are you talking about? Military spending creates new jobs, and the technologies created by military R&D eventually trickles down to the civilian sector. Look at high-impact plastics, computers, TCP/IP, etc.
You also mention that Japan and Germany are large economic powers, but that has a fair bit to do with the fact that they were both important, strategically, to American interests. So much so that the United States bankrolled almost the entire economic recovery in many parts of Western Europe, Germany included, and Japan. Japanese economy didn't start to recover from postwar periods until the Korean War, when Japanese factories were used to repair damaged American fighters.
Military spending becomes a liability when the flood of money stops. Eisenhower warned that *if* the United States creates an entire industry solely relegated to defense (as we did and still do), *when* the monetary flow slowed to that industry, *then* the economy will suffer. However, military R&D by itself will *not* be an economic liability, unless accompanied by shortsighted spending policies (as evidenced in the Reagan + Bush I years).