Slashdot Mirror

Top500 classifies supercomputers as clusters, constellations, and MPPs (Massively Parallel Processors). BG/L is an MPP. It is obviously not a cluster since a single compute node is not a general-purpose machine. BG/L is specifically intended for massively parallel applications.

The difference between clusters and constellations is the ratio between the number of nodes (that can either work independently or be clustered) and the number of processors in a node.

Go to the Sublist generator of Top500 to see what machines belong to which category.

See, e.g., definitions here

Top500 classifies supercomputers as clusters, constellations, and MPPs (Massively Parallel Processors). BG/L is an MPP. It is obviously not a cluster since a single compute node is not a general-purpose machine. BG/L is specifically intended for massively parallel applications.

The difference between clusters and constellations is the ratio between the number of nodes (that can either work independently or be clustered) and the number of processors in a node.

Go to the Sublist generator of Top500 to see what machines belong to which category.

See, e.g., definitions here

What about

#15 MACH5 at COLSA http://www.top500.org/system/7741
#85 Turing at University of Illinois http://www.top500.org/system/7366
#308 Xseed at Bowie State University http://www.top500.org/system/7500

What about

#15 MACH5 at COLSA http://www.top500.org/system/7741
#85 Turing at University of Illinois http://www.top500.org/system/7366
#308 Xseed at Bowie State University http://www.top500.org/system/7500

What about

#15 MACH5 at COLSA http://www.top500.org/system/7741
#85 Turing at University of Illinois http://www.top500.org/system/7366
#308 Xseed at Bowie State University http://www.top500.org/system/7500

These days about half the supercomputers simply use Gigabit Ethernet top500 site. And yes, that's commodity. Dual Gig Ethernet ports can be found on even $700 off the shelf servers these days.

Worrying abou topology is extremely importing. Spending your money on $1500/node interconnects probably isn't. The highest-performance guys certainly don't use such mid-range interconnects as you described (that perform about like gigE), but reather build their own proprietary interconnects (the dark blue on the chart I linked).

You seem to misread the current trend: Amd has been a no-go for a long time - due to the lack of decent mainboards mainly. It's increasing its share. Intel has lots of systems because of the system they've already installed - not because they're more of them right now.

Facts? November 2004:
Intel 63.6%, AMD 6.2%.

November 2005:
Intel 66.6%, AMD 11%

IOW: AMD almost doubles its share, wins 4.8%; Intel wins 3%

You seem to misread the current trend: Amd has been a no-go for a long time - due to the lack of decent mainboards mainly. It's increasing its share. Intel has lots of systems because of the system they've already installed - not because they're more of them right now.

Facts? November 2004:
Intel 63.6%, AMD 6.2%.

November 2005:
Intel 66.6%, AMD 11%

IOW: AMD almost doubles its share, wins 4.8%; Intel wins 3%

Click on the name of the supercomputer and then click on "System Details". The OS is listed there. Like so: http://www.top500.org/system/details/7747

could withstand the mighty power of a Slashdotting!

ERROR
The requested URL could not be retrieved

While trying to retrieve the URL: http://www.top500.org/lists/2005/11/basic

The following error was encountered:
        * Connection Failed

The system returned:
        (60) Connection timed out

The remote host or network may be down. Please try the request again.

Looking at this chart http://www.top500.org/lists/2005/11/l/Operating_Sy stem it actually appears that the OS ran on all system are:
- Linux: 72.2%
- Max OS: 1.0%
- Others 4.4%
- UNIX and Linux: everything else (~22%)

So it appears that Linux/UNIX* runs on about 95% of all super computers. The Story headline should have been:
Linux Claims Almost All Supercomputers Spots

What a scoop.

*Linux,UNIX, what's the difference really?

But our supercomputer, System X is the top supercomputer in academia! :-D (Man, it's already been slashdotted!) -Acercanto You can have only two of the following three qualities when developing a product: cheap, fast or good. You can produce something cheap and fast, but it won't be good, good and fast, but it won't be cheap, good and cheap, but it won't be fast.

...and x86 is faster.

Oh, you poor, uninformed miscreant.

Instruction set has nothing to do with speed. There might be faster x86 processors available, but there's also fast IA64 processors, PPC processors, x86_64 processors, etc.

At best, x86 is slower than RISC based processors since, without fancy scheduling, and other nifty tricks, RISC processors do more per cycle than their x86 counterparts. See the, albeit flawed, benchmarks comparing G5s to P4s. Better yet, look at the top 500 supercomputers notice how #1 BlueGene/L is running on PowerPC 440s. Not until #10 do you get an x86 based architecture (even then it's x86_64).

icc does a shithouse job compiling ATLAS. If you're doing linear algebra and using BLAS libraries (doing, oh, I don't know high performance computing), gcc outperforms icc by 40-50% on Opterons. Go figure. gcc compiles the hand-tuned assembly, and icc skips it and compiles optimized C only, missing out on SSE and other enhancements.

The difference is no longer an order of magnitude or two even for the SMP boxes that you are talking about. Googling a bit, I have found these reasonably priced Opteron boxes. These are variations of the same opteron v40z boxes that Sun sells. Quad Opteron boxes w/ hot swap power supplies and hot swap hard drives w/ ECC memory for reasonably cheap for the chassis.

I've never seen a terabyte of ram on a single domain but the
very high end sun boxes will allow you to use either 576 or 288 gigs of ram on a single box, so I will use that as the base. The maximum in SMP mode for the opteron boxes you can homebrew is probably one based on the iwill motherboard which supports 64 gigs, which is now grabbing the midrange line of the sun server series.

In fact HP makes the ProLiant DL585 with opteron chips. They will sell you this machine w/ up to 128 gigabytes of ram, redundant roms, redundant power supplies, hot swappable hard drives, ECC memory.

Its not quite at the high end server sun range but the opteron chip and its use in machines built by traditional "big iron" builders is not really an order or magnitude off. Its really about a factor of 5 from the really high end (aka $1m+ sunfire installations).

And as the IWILL MB points out its not even that far away from home brew or at the very least a $100k officebrew given the amount of ram one has to buy.

Also its worth pointing out the processor supports up to a terabyte, the issue is the memory sticks/mb's are not available yet.

And here we are only talking about the SMP/redundancy style computing . For distributed style computing platforms (aka everything on the top 500 super computers), the 10th and 11th fastest computers in the world are Cray machines based on opterons currently running at Sandia National Labs and Oakridge National Labs respectively.

Now perhaps in your world the top 500 super computers are not considered big iron because they are distributed but i'm w/ the joe slashdotter crowd that considers "big iron" to include the top 500 fastest computers in the world.

The difference is no longer an order of magnitude or two even for the SMP boxes that you are talking about. Googling a bit, I have found these reasonably priced Opteron boxes. These are variations of the same opteron v40z boxes that Sun sells. Quad Opteron boxes w/ hot swap power supplies and hot swap hard drives w/ ECC memory for reasonably cheap for the chassis.

I've never seen a terabyte of ram on a single domain but the
very high end sun boxes will allow you to use either 576 or 288 gigs of ram on a single box, so I will use that as the base. The maximum in SMP mode for the opteron boxes you can homebrew is probably one based on the iwill motherboard which supports 64 gigs, which is now grabbing the midrange line of the sun server series.

In fact HP makes the ProLiant DL585 with opteron chips. They will sell you this machine w/ up to 128 gigabytes of ram, redundant roms, redundant power supplies, hot swappable hard drives, ECC memory.

Its not quite at the high end server sun range but the opteron chip and its use in machines built by traditional "big iron" builders is not really an order or magnitude off. Its really about a factor of 5 from the really high end (aka $1m+ sunfire installations).

And as the IWILL MB points out its not even that far away from home brew or at the very least a $100k officebrew given the amount of ram one has to buy.

Also its worth pointing out the processor supports up to a terabyte, the issue is the memory sticks/mb's are not available yet.

And here we are only talking about the SMP/redundancy style computing . For distributed style computing platforms (aka everything on the top 500 super computers), the 10th and 11th fastest computers in the world are Cray machines based on opterons currently running at Sandia National Labs and Oakridge National Labs respectively.

Now perhaps in your world the top 500 super computers are not considered big iron because they are distributed but i'm w/ the joe slashdotter crowd that considers "big iron" to include the top 500 fastest computers in the world.

There is a Cray XT3 that runs at 15 Teraflops at Sandia and made out of 2ghz opterons and is currently the 10th fastest computer in the world. There is a similar machine over at Oak Ridge National Labs that runs at 14 Teraflops and is the 11th fastest computer in the world.

In fact, those lowly AMD kids seem to also have their chips on the fastest machine at the Pittsburgh supercomputing center (ranked 33rd fastest computer in the world) and the US Army Research Laboratory (ranked 39th fastest) . The latter was actually being built by IBM for ARL, you know those guys who coined the term "big iron".

There is a Cray XT3 that runs at 15 Teraflops at Sandia and made out of 2ghz opterons and is currently the 10th fastest computer in the world. There is a similar machine over at Oak Ridge National Labs that runs at 14 Teraflops and is the 11th fastest computer in the world.

In fact, those lowly AMD kids seem to also have their chips on the fastest machine at the Pittsburgh supercomputing center (ranked 33rd fastest computer in the world) and the US Army Research Laboratory (ranked 39th fastest) . The latter was actually being built by IBM for ARL, you know those guys who coined the term "big iron".

There is a Cray XT3 that runs at 15 Teraflops at Sandia and made out of 2ghz opterons and is currently the 10th fastest computer in the world. There is a similar machine over at Oak Ridge National Labs that runs at 14 Teraflops and is the 11th fastest computer in the world.

In fact, those lowly AMD kids seem to also have their chips on the fastest machine at the Pittsburgh supercomputing center (ranked 33rd fastest computer in the world) and the US Army Research Laboratory (ranked 39th fastest) . The latter was actually being built by IBM for ARL, you know those guys who coined the term "big iron".

There is a Cray XT3 that runs at 15 Teraflops at Sandia and made out of 2ghz opterons and is currently the 10th fastest computer in the world. There is a similar machine over at Oak Ridge National Labs that runs at 14 Teraflops and is the 11th fastest computer in the world.

In fact, those lowly AMD kids seem to also have their chips on the fastest machine at the Pittsburgh supercomputing center (ranked 33rd fastest computer in the world) and the US Army Research Laboratory (ranked 39th fastest) . The latter was actually being built by IBM for ARL, you know those guys who coined the term "big iron".

China has had that system running since June 2004: http://www.top500.org/sublist/System.php?id=7036
Which makes my wonder, how much of the post is really news?

Benchmarks! Wooo! So now we know that Java can run benchmarks and example programs really fast.

Actually, those aren't just benchmarks. They are code that has been used for decades because they illustrate real use. The numerical methods they test are used in thousands of applications that require high-performance numerics and data structure handling.

That is why Linpack isn't just any old benchmark - it is respected. It is used to rank the performance of supercomputers.

http://www.top500.org/lists/linpack.php

That Java is up to running Linpack at speed is a significant and unquestionable indication of high performance.

Now let's look at the real world instead. Why not download Azureus or AEJEE and compare them to similar apps that are written in native code?

I have tried AEJEE. I have to admit it is pretty awful! The GUI is slow. It looks like some sort of dreadful GNOME emulation.

However, why not try another real-world app - JEdit. Under Java 1.5 it starts up fast and has a GUI speed that no-one can complain about.

I use Java GUI apps (NetBeans) for my development. I would be the first to say that it was slow and barely usable many years ago, but now it is fast and sleek.

I would be interested in what you mean by native code? C? C++? They are translated to native code before they run. Java? That is native code - translated to native code as it runs (this cached native code is then re-used). I guess you must mean hand-coded assembler... well who does that these days?

Not just misleading on mpp/cluster. Also dated with respect to Amdahl's law.

William Camp, a center director (http://www.cs.sandia.gov/), from Sandia National Laboratories, says that "contrary to Amdahl's expectation, we can routinely" do better. He also says that "Amdahl neglected the overhead due to communications". From a talk found at http://www.blu.org/meetings/2003/11/NovBlu.pdf and qualified a little more in http://www.top500.org/corner/articles/article_15.p hp.

DoE currently has a 136TFLOP cluster . They use it for more than nuclear weapons research.

IBM can most certainly do low voltage (and low power) PowerPCs -- just look at the BlueGene compute chip, for example.
Even the Cell is considered relatively low power at low voltages, especially considering it's a 9 core uP.

Apple's decision was about cost, plain and simple.
They were too cheap to prod IBM into being interested in spending money on developing another desktop processor, especially when it must have been difficult in keeping IBM's interest what with scoring the entire game console market.

Woah now, I think bitterness is an understatement.

Supermike is, in fact, fully operational. I beleive being operational is a key requirement in being listed in the top500.org list

I think the results speak for themselves.
2003 ACM Regional Programming contest results:
LSU Teams ranked: 21st, 27th, 50th, 67th, and 72nd
ULL Teams ranked: 70th

2002 ACM Regional Programming contest results:
LSU: 4th, 17th, 37th, 57, 63
ULL: 18th, 46th

Perhaps if ULL isn't getting the "money and the press" that LSU is getting, you should take it up with the heads of the CSC department.

But in closing, I think the greatest evidence against LSU's computer science program being just a pony show is that LSU doesn't even HAVE an accredited computer science program! (hint: ULL does)

http://top500.org/sublist/System.php?id=6084
http://icpc.baylor.edu/icpc/regionals/ViewRegional Standings.asp?ContestID=648
http://acm2002.csc.lsu.edu/results/finals.html

National Center for Supercomputing Applications, a research branch of the University of Illinois at Urbana-Champaign.

Their claims to fame include having four supercomputers of the 50 fastest in the world, and creating Mosaic, the first graphical web browser.

If you have doubts about the influence of Mosaic, load up internet explorer, click "Help" in the menu, then click "About Internet Explorer" and read the blurb....

Calculating pi is a series of mathematical operations where you can't do the next one without the prior because you need the remainders. Supercomputers are super due to a heck of a lot of CPUs all working on different parts of a problem that can be broken into chunks. How exactly do you break a series of operations that depend on the priors into chunks for a supercomputer to rip through?

Use the BBP Formula. Pifast is just a benchark, like all benchmarks it's rather silly. The record is for PCs, the top 500 supercomputers are benchmarked using another silly benchmark (LINPACK).

From what I recall about Peloton(that's what the presenter called it), they wish to have a 14.8 TF/s scalable unit with 4x Infiniband interconnect. This scalable unit itself is more than half the power of Thunder(ranked 7 in Top 500) http://top500.org/lists/plists.php?Y=2005&M=06 They plan to have 16 such scalable units.

For those who are interested in the specs: Peloton is 16 SU with 236.5 TeraFLOP/s, 215 TiB memory, 5.0 PB global disk system with 6,720 SMPs and 48+24 = 72 IBA 4x DDR sw. Power is 4.05 MW.

but only if you live in Fanboy Land where Linux is the answer to every computer-related problem.

Or, for instance, you run the fastest supercomputer in the world.

Cheers
Stor

One could also argue that Linux is what FreeBSD should have been, and cite the huge number of supercomputers using Linux, or the success of Linux on the mainframe. However, it would be nice if the poster realized that it's a pissing contest and both operating systems are impressive and have their uses, benefits, and drawbacks. Neither is what one "should have been". They both have their own, very different methodologies, so let's leave it at that.

Not that it's news anyways...

I believe the most powerful supercomputer in Europe is in Spain rather than the Netherlands as mentioned on th BBC website.

You are right, according to this list the Barcelona Supercomputer is slightly faster. This hurts my dutch pride.

The same goes for the "real", not HPC servers, which run on Power4+ or Power5, which also has nearly nothing in common with those chips.

Where's the market....

http://top500.org/sublist/System.php?id=7605

I for one can't wait until OSS solutions provide the most popular webserver or run on the fastest computer.

Obviously you wouldn't want to run Linux on a massive 64 processor server

Actually sir... that's not... entirely... true.

In fact, the historical performance of supercomputers is one of the strongest proofs that the Moore's Law will continue unabated forever (*)

http://www.top500.org/lists/2004/11/perfdevel-spec ial/projected.php

The total performance of the 500 fastest supercomputers has been increasing almost perfectly exponentially from 1993. The graph is very impressive.

* Insert standard disclaimer about forward-looking statements here.

They let you make pretty tables

To answer your question and clear up the incorrect info posted:
#systems, %, sum Rmax, sum Rpeak, #processors
Intel+AMD = 358 71.6 895150 1480709 266242
Power+PC = 77 15.4 616413 897521 251502
All = 500 100 % 1686933 2632133 580336
%Rmax, %Rpeak, %processors
Intel+AMD = 53.1% 56.3% 45.9%
Power+PC = 36.5% 34.1% 43.3%

On a side note, how can I format this better, I've tried some html tags but slashdot ignored them.

http://www.top500.org/sublist/System.php?id=6560

Cornell is using a Windows cluster. It is ranked 326.

top500.org provided statistics for all top 500. For breakdown by processor family, see http://top500.org/sublist/stats/index.php?list=25& type=procfam&submit=Generate+Table. Assuming your summary for top 50 is correct, the statistics is rather different from top 500.

I don't know why top500.org didn't provide breakdown by operating system, so I found out myself. Here it is:

328 (65.6%): Linux
73 (14.6%): HP Unix (HP-UX)
52 (10.4%): AIX
16 (3.2%): UNICOS
7 (1.4%): Super-UX
6 (1.2%): Solaris
4 (0.8%): Tru64 UNIX
4 (0.8%): MacOS X
3 (0.6%): SuSE Linux Enterprise Server 9
2 (0.4%): Redhat Enterprise 3
2 (0.4%): HI-UX/MPP
1 (0.2%): SuSE Linux Enterprise Server 8
1 (0.2%): Paragon OS
1 (0.2%): IRIX

I expected a few Windows, but surprisingly there is none at all. Not sure how accurate top500.org's "Operating System" field value is though.

I work in Oak Ridge, TN, so I'm at least near to four Top 500 systems. That's as good as it gets for me.

The Top500 site says that machine is running Linux though that may be a mistake.

Earth Simulator Facts
BlueGene/L Facts

Earth Simulator Facts
BlueGene/L Facts

...and the link to the Top500 list sorted by operating system -- which tells the story of the rise in Linux.

Molecular dynamics simulations.

Actually, there is one, the Cornell Theory Center, ranked at #326.

Does anyone else think it's wierd that a graphic design company like Animal Logic has 2 of the top 500 supercomputers? Or whatever they call Gaming Company B in China, with the 2.2TFLOPS at each of #150, 151 & 153, or their Taiwanese counterpart at #152?

It seems that M$ doesn't enter into this picture. But it's interesting to see how Linux is grownig.

Wrong, Alpha is still fourth. With half of the processors of the 3rd competitor (5120 vs 10000) it achieves more than half of its "R peak". I guess there're lot of factors (interconnexion use, etc) here but it'd be interesting to see a performance/number of CPUs chart