Slashdot Mirror
Top 500 Supercomputers
Anonymous Coward writes "sendmail.net has a piece on the new Top500 list of supercomputers. 'So who came out on top? Well, three US Department of Energy machines have taken spots one, two, and three to lead the list: ASCI Red (manufactured by Intel) at Sandia National Labs in Albuquerque, ASCI Blue-Pacific (IBM) at Lawrence Livermore Labs in Berkeley, and ASCI Blue Mountain (SGI) at Los Alamos. These are the only three systems to exceed 1 TF/s on the Linpack benchmark, and represent 7.4 percent of the total Flop/s on the list.' The story notes that the average growth rates for the list exceed the number set by Moore's Law. "
A better metric for computation power is given by this formula concerning the memory heirarchy:
bandwidth*size*speed where memory
bandwidth is the average speed at which data streams to or from memory
size is the amount of memory
speed is the responsiveness to random access
What the massively parallel processor advocates frequently forget is that locality of reference is an expensive assumption. A similar mistake is made by memory heirarchy advocates. For example, many systems where CD-ROM jukeboxes were included to expand the size of the memory the architects overestimated "locality of reference" and therefore underestimated the profound impact that moving the robot arm around would have on latency. Such designs are convenient for the hardward designer who wants "good numbers" and a nighmare for advanced software application that needs unpredictable access to lots of information at a high rate in order to get the solution out of the machine before the solution is obsolete. The operands have to come together through that maze of wiring. If you have partitioned the memory, it profoundly affects both latency and bandwidth. The critical thing is to allow _shared memory_, and that means advanced memory control units.
Seymour Cray kept ahead of the supercomputer pack for more than two decades by focusing his best talent on fast, high bandwidth memory control units and building the biggest semiconductor memories to match.
Seastead this.
That doesn't exactly hold. Replace my use of "speed" with "transistors" or "density", and it is still the same. Every 18-24 months density/transistors/speed may double in chips /on average/, but at the same time there are very expensive ones with higher densities/whatever that are faster, and very cheap ones with lower densities/whatever, so to remove the money factor is not fair. At any point in time, regardless of moore's law, anyone can build a faster machine simply by throwing more money, more resources at it. So it is not fair to compare machines like this. They should be normalized to density or performance per dollar (or any given monetary unit).
It's 10 PM. Do you know if you're un-American?
Right...Moore's law won't break when you add more hardware. The density will still be the same. But the cost won't. These systems should be compared on equal footing. If my $5000 system runs at 70% of the speed of your $10000 system, I win, because I have a better performance to cost ratio. Theoretically, add another of my $5000 systems, and my composite $10000 system beats your $10000 system. That's how things should be compared. It makes no sense to say my $X K7 beats your $.5X Celeron. It's bang for the buck that matters (more bang per buck indicating better design).
It's 10 PM. Do you know if you're un-American?
Massively parallel machines tend not to "run" an OS as much as they are "run by" an OS. The control nodes or stations (often separate machines) run whatever (i.e. Unix) and the actual compute nodes run their computation and little else. After all, running 10000 copies of Linux isn't really the best use of resources.
Drinking will help us plan!
The Sandia/Intel ASCI-red TFLOPS machine has proven to be one of the more technically successful efforts in massively parallel, high-performance computing. However, large MPP systems have drawbacks. Among these are:
Applications that require high levels of compute performance will continue to grow in size, variety, and complexity. While cluster-based projects have firmly established a foundation upon which small- and medium-scale clusters can be based, the current state of cluster technology does not support scaling to the level of compute performance, usability, and reliability of large MPP systems. In contrast, large-scale MPP systems have addressed the problems related to scalability, but are limited by their use of custom components. In order to scale clusters to thousands of nodes, the following must be addressed: Use of non-scalable technology must be bounded or eliminated. Technologies like TCP/IP, NFS, and rsh have inherent scalability limitations. Scalable management and maintenance is critical. The complexity of maintaining the cluster should not increase as it grows. Usability of the machine is critical. Users should not be required to know detailed information about the cluster, such as the name of each node or which nodes are operational, to effectively use the machine.
-----------------------------------------------
The IBM SP series machines I've run into all ran unix.
The Suns of course run unix.
I would guess the HP's run unix, although it might not be HP-UX.
The SGIs probably run IRIX unless they are "Cray/SGI" T3Es in which case they run Unicos
I don't know about the NEC or Fujitsu machines.
Just a nit...
Livermore NL is in Livermore. Berkeley NL is the one in Berkeley.
I teach Unix courses (on Linux) and in the first class I try to give an idea of where Unix is used.
I always say "the fastest computers in the world run Unix", but I'd rather be able to say "480 of the top 500 computers run Unix" - it sounds more impressive. The problem is that, although I can identify most of the operating systems on the list quite easily, I'm not sure about some of the more esoteric ones. Does anybody know exactly what all these systems are running?
It is tempting, if the only tool you have is a hammer, to treat everything as if it were a nail. - Abraham Maslow
There's a user manual available here for ASCI Blue. LLNL is already working on a 10 teraOPS machine called ASCI White. 8000 processors... ASCI Red is currently 1.8 TeraOPS.
Augh. Should have used preview. Should be:
170 NEC NLR 8 - fastest computer with a number of processors less than 10
101 SGI "Government" 1024 - PRESUMED slowest computer with a number of processors greater than 1000
teach me to consider a less-than symbol "Plain Old Text."
-=Best Viewed Using [INLINE]=-
...Beowulf cluster you could make out of those!
(Score: -1, Unoriginal)
Seriously, all we really want to know is which of the machines on the list are Linux clusters of some sort. This is still Slashdot, after all...
--
Xenu loves you!
Yes, the law should more accurately be stated "Speed /per doller/ doubles every 18-24 months"
Throwing money at the problem isn't fair...perhaps they should have normalized these systems based on their price...
It's 10 PM. Do you know if you're un-American?
Didn't we hear something a few months back about a 1000-node cluster of Alphas doing genetic algorithms? I wonder how that would have measured up.
Also, it would be nice if the table added a column for Rmax/AcquisitionCost, and let you sort by that column. I'll bet that would launch some Beowulfen toward the top.
--
It's October 6th. Where's W2K? Over the horizon again, eh?
Sheesh, evil *and* a jerk. -- Jade
Yes, they're on there. They're called 'Self Made'. Here are a few:
#44 CPlant Cluster
#265 Avalog Cluster
#454 Parnass2 Cluster
-- I'm the root of all that's evil, but you can call me cookie..
The Hitachi machine can achieve these figures for two reasons:
1) Their Interconnect
2) Their Processors
The interconnect is a hyper-bar crossbar network, with a bandwidth of 1GByte. Also they are able to get sustained message passing performance of about 90% like they did on their previous machine the SR2201. Other vendors would provide 60-65% of peak.
The number listed in the Top500 for processors is a bit mis-leading, this is in fact the number of nodes. The Hitachi nodes are made up of a number of processors, each with pseudo-vector optimisation (allowing them to miss the cache when loading large memory blocks). This optimisation means the chip can have a high sustained performance on large scale numeric problems. The nodes can be configured as either SMP of vector. This allows the machine to address a much wider range of domain problems.
Hitachi have a very brief page describing their machines SR8000 Product Page
I would love to see what a fully configured machine could do (6 TFlops!).
BTW, Linpack is not a great gauge of a Supercomputers performance. When there a lot of nodes it becomes message bound and does not reflect the true performance of the machine. When looking at machines like this it is important to look at benchmarks related to domain problems. e.g. It does not really matter what interconnect you have if you are doing ray-tracing, but it matters a great deal when doing astro-physics.
Does Moore's Law apply to massively parallel, hand-built systems?
I wouldn't think so... At the time, Moore was running Intel, a one-CPU-per-machine outfit, and I think his "law" was an observation on the rate of progress in the PC industry, and what advancement was possible within the technology of single Von Neumann-bottleneck-style systems.
-schmaltz
Big Daddy, Johnny, Burp, Aunt Zelda, Scott, Slurp, Big Momma
I hate articles like this. In the first place theres the, "The new Top500 numbers are in, and your laptop has never looked so tragically slow." These supercomputers are all massively parallelized machines using regular microprocessors. The actual speed of the machine, like ASCI Red, is determined by the processors used, which in this case are just normal Intel processors. So you can go out and buy a machine that computes instructions just as fast as ASCI Red. The difference is that it can do more things at once because of all the processors involved. Does it make your laptop look slow? Hell no, because if you had ASCI Red, you wouldn't have any apps that take advantage of its parallelism to run on it anyway.
r e.htm):
Secondly, Moore's Law is the following (from http://www.intel.com/intel/museum/25anniv/hof/moo
In 1965, Gordon Moore was preparing a speech and made a
memorable observation. When he started to graph data about the
growth in memory chip performance, he realized there was a striking
trend. Each new chip contained roughly twice as much capacity as its
predecessor, and each chip was released within 18-24 months of the
previous chip. If this trend continued, he reasoned, computing power
would rise exponentially over relatively brief periods of time.
Moore's observation, now known as Moore's Law, described a trend
that has continued and is still remarkably accurate. It is the basis for
many planners' performance forecasts. In 26 years the number of
transistors on a chip has increased more than 3,200 times, from 2,300
on the 4004 in 1971 to 7.5 million on the Pentium® II processor.
Since the CPUs in supercomputers use standard processors, and Moore's Law applies to these processors, his law is still intact. His law is about CPUs, not systems.
He said, "You'll be able to tell your grandchildren that you helped assemble the first NT supercomputer," and I cringed.
Ishtar (1 flop)
Kevin Costner (1 megaflop/year)
--
"L'IT c'est moi!"
The real action is lower down Avalon was top 100 now is down to 265. The top cluster goes to cplant take the award for top cluster now.
Grey (Chris Lusena)
The only reason this term was thrown around is cause in a test, the G4 exceeded the 'allowed exportable power' for IC's. Personally, that's not enough to make me switch to the Apple market. I'm happy with my slow, but configurable system.
NIVRAM
94 IBM MHPCC 243 - fastest computer with a number of processors in no way related to common powers of 2
hmm, 243 is 3^5. I wonder what strange architecture dictated that number.
No. The sordid truth is that the Republicans are marginally behind the Democrats and the Pentagon, in the Mega Deathmatch, currently being played on a network of supercomputers and an enhanced Quake server.
The Republicans are desperate not to lose precious cycles to simulations, which would give the other two teams a decisive advantage. The lag might even cost them the tournament.
These would be OK at cracking DES, but really, given that DES can be cracked in less than a day on a kit computer (and within a week by assembling kitchen equiptment), DES is essentially dead, as far as the US Government is concerned. Only people like NASA and Boeing still use DES, to any degree. There are FAR worse vulnerabilities to their approaches, though, than mere crackability. Key management is - to be blunt - pathetic.
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
10 IBM UCSD 960 - fastest computer with a number of processors evenly divisible by 10
12 IBM Charles Schwab 2000 - fastest computer with a number of processors evenly divisible by 100
15 Fujitsu Kyoto 63 - fastest computer with a number of processors not evenly divisible by 2
46 Fujitsu NAL 167 - fastest computer with a number of processors neither evenly divisible by 2 nor equal to (2^x)-1
94 IBM MHPCC 243 - fastest computer with a number of processors in no way related to common powers of 2
170 NEC NLR 8 - fastest computer with a number of processors 1000
(Yeah, yeah, I know you've got 2k TRS-80s in a Beowulf cluster in your back yard.)
-=Best Viewed Using [INLINE]=-
I mean, it's nice to see the US machines taking the cream of the honors in raw power... but what the heck - ASCI Red gets it's 1st place berth with 9 THOUSAND some odd cpus (.0246Rmax pts / cpu), whereas the Hitachi machine gets a very respectable 5th with only 128: 6.8Rmax pts per CPU! Isn't there some credit due for the more efficient machine? It doesn't seem that impressive to simply dump silicon at a problem until you are #1...
Isn't it also true, however, that there are practical limitations of how many components you can add and still expect a reasonable increase in performance? i.e. After 512 processors it is no longer cost effective to add processors because of other limitations. I don't know the exact numbers. But then wouldn't moore's law apply along the limitations of these machines? Don't quote me on this, I'm just making it up.- -------------
--------------------------------------------
Yes, I do see a lot of entries from Intel, which means to me probably Pentium Pro's.
Then there's IBM, which seems to be using PowerPC 604e's.
Next, SGI uses MIPS
SGI/Cray - have they moved to MIPS, or are they still using Alpha's? (I'm not 100% sure that's what they used before, but i'm 95% sure it is).
My main puzzler here is NEC. WHAT ARE THEY USING??? If you go down to #73 on the list, there's a machine that was deployed in 1999 with just 16 processors? Okay, it's performance is 1/19th that of Intel's #1 offering, but it uses just 1/602 the amount of CPU's??? That's not a standard processor that i've ever heard of?
NEC has a bunch of listing below that, too. Some use just 5 processors (though, those are all in the high 400's). What chips is it using? Can anyone explain what this machine is?
I think it is very interesting to note that of the top 40 supercomputers, there are seven with less than 400 processors and every single one of them is located outside the United States (six in Japan, one in France).