Slashdot Mirror
Inside The World's Most Advanced Computer
Junky191 writes: "Just came across an informational page for the Earth Simulator computer, which provides nice graphics of the layout of the machine and its support structure, as well as details about exactly what types of problems it solves. Fascinating for the engineering problems tackled- how would you organize a 5,120 processor system capable of 40Tflops, and of course don't forget about the 10TB of shared memory." Take note -- donour writes: "well, the new list of supercomputer rankings is up today. I have to say that the Earth Simulator is quite impressive, from both a performance and architectural standpoint."
"Earth Simulator" is a rather bold name for a supercomputer, especially when you consider it probably can't even simulate the global weather fast enough to predict it (or even tell you what the weather is in real time). The computer looks impressive, but I think they should have stuck to a more abstract name rather than what I see as false advertisement.
if you could install the clients on this thing, you could find the cure to cancer, crack RC5-64 and OGR-25, decipher all of the SETI@Home work units (but you still wouldn't find any aliens :)...hell, you could solve ALL of the distributed computing applications on this machine.
:)
Stating the obvious, that's a shitload of CPU power.
I am not going to ask "Does this run Linux ?" because it obviously does not, but can anyone point to some good resources on what kind of Operating Systems do these monster machines run ? Are they some kind of a UNIX ? Or are they some elite breed of OS that mortal humans have no chance of understanding ? Linkage appreciated.
What is really amazing is that in 50-60 years, this amount of computing power will easily fit within the confines of the standard PC case (assuming such a thing even exists 50-60 years from now). Remember ENIAC...
*Condense fact from the vapor of nuance*
This is a question, and not a statement
While this does a nice job of crunching numbers, how do they know that their algorithms are any good at doing what they do? Or are they trying to simulate things that aren't continuously kicked around by chaos theory?
I ask because I've been looking at dynamics in my spare time, and simulating something as small as cigarette smoke accurately seems impossible (although I must say Jos Stam and Co did a nice job of making it look real). So it seems a bit bewildering to see something trying to simulate the earth, even if only at a macro level.
Seems to me that though ES takes the overall performance crown, that the IBM and HP (man that sounds strange) units have some definite advantages over it. Primary of which is the fact that they DO use "off the shelf" parts. ASCI White uses 375Mhz Power3 chips which are comparitively low performance compared to what IBM is shipping now (1.3 Ghz Power4). I don't know what the technical details are behind ASCI White, but it seems that IBM could instantly get a doubling of performance by using new CPU modules. With the "specialized processor" approach that NEC uses, this would seem to be prohibitively expensive. IBM has already amortized most of the cost of the development of new processors through their normal business units.
Another advantage would be that since ASCI White is a hyper RS6K, you could use a lower end model (and IBM could rather inexpensively offer a lower end model) to develop your models on before using the relatively expensive big boy to do the actual simulations. I have to admit that this point is moot if they don't keep the utilization of the thing up pretty high most of the time.
Though they mention that ES "only needs 5104" processors vs 8192 for AW, it looks like ES still takes up massive amounts of space. Now ES' storage is significantly larger that AW, so maybe that's where all the space is being eaten, but it would be interesting to see what the actual cabinet space/power requirements for the two machines sans storage are (assuming they are both using standard stuff for storage).
Others things include since AW is based on OTS parts, is it easier to get parts for when processing units konk out. Is it simpler for a tech to work on the unit. Since Linux is already running on RS6K, theoretically with a few device drivers, you could run Linux on that bad boy
Of course all this is moot in the non-real-world of supercomputers. With seemingly infinite budgets, the only _real_ measure is absolute performance, and ES obviously has the edge here. But if I were the IBM sales rep for supercomputing, I'd sure be hyping the fact that when it's not simulating nuclear explosions, you can run Gimp and Mozilla.
Are there any estimates of the processing power of all the worldwide computers participating in the SETI project?
Specialization? What specialization? The Top500 rankings are based on LinPack-- a software package for solving dense systems of linear equations-- which seems applicable to a fairly general set of scientific problems.
Do supercomputing manufacturers cheat on benchmarks? I don't know. Presumably it would be a rather expensive proposition-- and since supercomputing sites will benchmark with a variety of specialized and general purpose libraries, it seems unlikely to work.
There, are, of course, differences between weather simulations and galactic evolution simulations. But field specific benchmarks are inappropriate for a site like Top500--the whole point of the site is to allow someone to analyse gross trends. "This memory architecture once dominated the rankings--now its used by only a few entries. Perhaps our next computer platform shouldn't be based on that architecture." (and possibly writing journal articles about it.)
In addition, general purpose supercomputing sites are relatively common.
Although they are setting up a quite cool Sun Fire Ultra Sparc Cluster running Solaris.
The setup will consist of 16 Sun Fire 6800 SMP nodes (1500 MHz, each node is a 24 processor SMP system with 24 GB shared main memory) and 4 Sun Fire 15K SMP nodes (1500 MHz, each having 72 processors and 144 GB of memain memory) giving an max. arithmetic performance of 4 TFlop/s.
Check the link to see for yourself (like you dont have anything better to do, right?).
Sad/funny part of the story: the cluster is going to be finished in 2003 ...
I should check Moores law on top 500 super computers...
Alt least know the world knows we do cool stuff too ...
There are some nice pictures on the ES site as well. I wonder if the colouration of the cabinets is there to prevent the engineers from getting lost..? :o)
Remember, there are no stupid questions. But there are a lot of inquisitive idiots.
Hans Moravec estimates it would take about 100 Trillion instructions per second to emulate the human brain. At 38 Tflops, Earth Simulator is in the ballpark. Maybe they should have called it human simulator, or just "Sim".