Top 500 Supercomputers Ranked

Shadow Wrought writes "The Register is reporting on (alternate ZDNet article) the latest list of the top 500 supercomputers in the world. Top of the list is the Earth Simulator Center in Yokohama, Japan, with a benchmark performance of 35.86 Tflop/s. HP and IBM claim 159 and 158 of the systems respectively. I wonder how many teraflops Deep Thought could have done?"

Comments

American Dominance in Supercomputers

[reporter]

Before anyone says that Japan is "ahead" of the USA again, let me quell any doubts about the superiority of American computer technology. The technology to build a HPC (high-performance computing) computer is identical to the technology to build a commercial computer. American companies have been increasingly focused on building commercial computers because that is where the profits are -- especially after the end of the cold war.

The NEC Earth Simulator is really just a different optimization point in the computer-design space. Huge amounts of bandwidth to memory and specialized vector-processing units tied to the processor core. The VLSI technology that NEC used to build these system is readily employed by Intel and IBM. So, if the latter companies wanted to build the world's fastest HPC computer, they could.

The 21st century is not PaxAsia. It is PaxAmericana. The hordes of immigrants flooding into this country to get the hell out of Asia should have been a big hint.

Re:American Dominance in Supercomputers

[agent dero]

The difference in manufacturing in Japan and the US are another extension of the cultural differences. (I believe)

For instance, in Japan many manufacturers are creating smaller devices, something that hits the market has to be smaller and more "efficent" (not costing lots of Yen in overhead) This is something typical of the Japanese culture. For example, you won't see some Japanese Guy driving around in an Escalade with his girlfriend, "just cause" You'd more likely find some Japanese Guy driving around in a Honda Hybrid car, cause it's cheaper to run than most cars, even though the price sucks.

Meanwhile, in the States, people want more POWER! They want the big ass SUV that crushes other cars, and small animals. The same goes with the computers. Something big and fast, regardless of the power consumption or general overhead of the machine.

Therefore, the target markets in the two countries are much different, so the products of the two will also be much different.

Some thoughts...

[anzha]

Interesting to note is that #3, #6, and #8 are all linux clusters. All three of which are at Livermore.

Cray's X1 also debuted, but it was much lower @112. However, it ought to be noted, that the examples out so far are only 60 processors at tops. As soon as the money gets ponied up, prolly at ORNL, they'll be waaaay up towards the top. My guess is, if all goes as planned, they'll be at #15 by year's end.

What I find exciting these days is actually the High Productivity Computing Systems Effort, the Blue Planet or Blue Gene. These are a little ways off from being on the Top500 list yet though. :D

I do wish there were more SC companies doing hardware development in the US. I love Cray, but a single vendor smacks of eggs in one backet syndrome...So, geeks, if ya wanna start a startup with a design, go for it...Betcha the NSA (aka Cthuhlu of HPC) would be happy to sponsor ya...;)

The top 500 unclassified supercomputers

[mikeophile]

What I want to see is the list amended to include the iron that agencies like the FBI, CIA, NSA, and less well known acronyms are using.

Re:The top 500 unclassified supercomputers

[hawkstone]

#2 (Q at LANL) and #4 (White at LLNL), and others, are doing classified work, but they're still on this list. But you're pretty much right if you meant that these are all unclassified architectures.

Interestingly, though, things the NSA has wouldn't be likely to show up on this list, as the benchmarks are suited towards MPP style machines. NSA is more likely to have vector machines than large numbers of scalar processors.

Re:The top 500 unclassified supercomputers

[sysjkb]

An alternative list of supercomputers is (or perhaps "was") kept by Gunter at GAPCON. He ranks sites, as opposed to individual systems. Unfortunatly, the site seems to be down right now, but the ever helpful google cache (search for "gapcon supercomputer") put the NSA at #4, behind the Japan Marine Science Center, LLNL and LANL.

12475.7 - (09-JAN-2002) [NSA]
National Security Agency,Fort Meade,Maryland,US
1) Cray X1-3/192 2457.6
2) Cray T3E-1200E LC1900 2280
3) Cray T3E-900 LC1324 1191.6
4) SGI 2800/250-2304 1152
5) HP SuperDome/552-512 1130.5
6) Cray T3E-1350 LC800 1080
7) SGI 3800/400-1064 851.2
8) Cray T3E-1200E LC540 648
9) Cray T3E-1200E LC540 648
10) Cray T3E-1200E LC540 648
11) Cray T3E-1200 LC404 484.8
12) Cray T3E-1200 LC284 340.8

I think much of his information comes out of press releases, leaks, submissions, etc. The record for the NSA hadn't been updated for a while, but this may give you an idea of what they might have been running not very long ago.

The modern stuff is all well and good, but what's really fascinating is his list of the top computing sites in 1956. (Or search google for 1956 computing sites and click on the cache.) Here's the top two. Rating is in OPS:

583733.3 - [ONR]
Office of Naval Research,Arlington,Virginia,US
1) MIT Whirlwind 1 500000
2) ERA Atlas 2 83333.3
3) ERA 1101 200
4) ERA Atlas 1 200

253787.8 - [MIT]
MIT,Cambridge,Massachusetts,US
1) MIT TX-0 166666.7
2) MIT Whirlwind 2 45454.4
3) IBM 704 41666.7

Note that the supreme super-computer of that era, the Whirlwind, is quite a bit slower than your pocket calculator.

Yours truly,
Jeffrey Boulier

What language?

[Aardpig]

I'd find it fascinating to see a breakdown of the languages used to write the applications which are running on these beasties. High Performance Computing has rather different needs from a language than programs that are, say, focused on interaction with a user, or database access. I expect that languages which sit well with infrastructures such as MPI and Open MP would be well-favoured, and I wouldn't be surprised to find that FORTRAN puts in a respectable showing.

And before I'm bombarded with comments pointing out how ancient FORTRAN is, it's worth remembering that FORTRAN is still an evolving language; the last updade came out in 1997/98, and the new FORTRAN 200x should be arriving within the next year or so. In my experience of a number of languages, I've found that FORTRAN still continues to excel at numerical efficiency and portability, and I hope these selling points continue to be a feature of the new standard. Of course, I wouldn't want to write a compiler in FORTRAN, but for stuff like computational fluid dynamics, it still rocks; and those aspects of FORTRAN 77 which made it awkward (such as lack of dynamic array allocation) are fortunately a thing of the distant past.

I nominate Pink

[DeathPenguin]

Pink is the largest LinuxBIOS cluster in the world to date. The only moving parts in each node are cooling fans. This translates into a small savings on hardware, but more importantly means the nodes are by design more reliable thanks to fewer moving parts. Each node is powered by two 2.4GHz Xeon processors with four gigs of RAM and Myrinet 2000 interconnects.

The raw hardware power, while impressive, isn't what makes this cluster unique. The kicker is in the software, more specifically Clustermatic 3 featuring LinuxBIOS. Stuff happens and nodes fail, but thanks to LinuxBIOS they can be back up in a matter of seconds, not minutes.

Additional tools for the frontend node from Linux Networx makes updating nodes super-easy. You can flash each node's BIOS with a single command all in a matter of seconds. BProc allows you to run basic shell commands on any node without even installing a distribution on those nodes. w00t!

What we see here is a big shift away from expensive hardware and proprietary software. The software powering this cluster is 100% GPL, so users save a fortune in software licensing costs alone. And while these P4 nodes in particular aren't exactly cheap, they provide pretty darn good power and are far less expensive than Alpha servers. Also, using the x86 architecture means that consumer boards are not far behind in clustering. In fact, you can check out the LinuxBIOS homepage and see some pretty cheap boards that are supported already. So if you have some spare cash lying around and a couple weekends to kill, you can pick up a cheap board + cpu + memory combo and set it up as a slave node for your desktop machine with the same software these guys use to power this huge cluster.