Latest Top 500 Supercomputer List Released

chrb writes "BBC News is reporting on the release of the June 2010 Top 500 Supercomputer list. Notable changes include a second Chinese supercomputer in the top ten. A graphical display enables viewing of the supercomputer list by speed, operating system, application, country, processor, and manufacturer."

Comments

June?!

[Anonymous Coward]

Holy crap, the supercomputers are so fast they're in the future!

Linux

[B5_geek]

Ya for Linux!

Seriously, if this doesn't make every PHB take notice I can't imagine what would. (Hey boss, its free too!)

Re:Linux

[Pharmboy]

All our admins and all of our users only know Microsoft systems. Training isn't free.

So your users can't use Linux on the server? Or is it that all the users use super computers on the desktop? Our biz has all MS on the desktop and all Linux on the server. Obviously it is completely seamless. As for the admins, any admin worth their salt is always learning new things to just keep up with technology as it changes. Learning Linux by installing it on one system to start is trivial, and in certain situations, much easier to setup than Windows, such as DNS servers, web servers, etc.

If your admins can only work on a server if it uses a mouse, you need new admins.

Re:Linux

[Black Art]

In my experience Windows admins require *MUCH* more training than Linux admins. There is much more "black magic" that they need to know to be good at their jobs.

A Windows admin needs to know all the secret registry hacks to make things run well. They need to know all the non-intuitive places that Microsoft hides the settings for whatever services need to be configured. They also need to know how to recover things when it all goes horribly wrong.

Most Linux systems have text files to configure things. The files are in a predictable place. Updates are pretty easy and clear.

But Microsoft has scammed people into believing that leaving it harder than just putting up with the same old crap. In this case I just wish that people did get what they pay for...

   

Re:Linux

[morgan_greywolf]

I've done systems administration on both platforms for years and I don't think that there is any real appreciable difference between the amount of knowledge and training needed on one vs. the other when comparing systems that perform similar functions. Compare Active Directory to OpenLDAP+Kerberos 5, for example. They are very, very similar in a lot of ways; so much so, in fact, that OpenLDAP+Kerberos 5 can be used to host the directory portion of a Windows domain.

Re:Linux

[evilviper]

I don't think that there is any real appreciable difference between the amount of knowledge and training needed on one vs. the other when comparing systems that perform similar functions.

I'm afraid I'd have to completely disagree with you there. My preferred example isn't Active Directory, but CIFS...

With Windows, you do all the user management, then click-through the painful server setup wizard to turn on file sharing, and everyone's happy and thinks it's oh so easy to do... Right up until some guy tries to upload a a big file, and for no reason everything hangs-up at about the 2GB mark.

Congratulations, you've been screwed by Microsoft. No way in hell will you be able to determine what's going wrong by examining the OS, logs, included help documentation, registry settings, etc. You are screwed. You must now find external sources of information to determine why this would possibly happen. After either buying some 3rd party books on the subject (Microsoft's books are crap), or scouring Microsoft's website for every possible keyword that might be in the description of this problem, you just might find the answer (or maybe not, it really does depend on luck).

The software from this the largest and most profitable company in the world, somehow consistently determines that your gigabit network is ACTUALLY a slow, high lag (most likely dial-up) link, and dynamically switched to some ancient version of the protocol to give you slightly less overhead, which seems to work fine at first blush, but just doesn't allow UPLOADING (downloading is okay) files over 2GBs. Armed with this knowledge, you now get to delve deep into the bowels of the registry, and find a half-dozen irrationally named keys and change some completely arbitrary DWORD values to some other completely arbitrary DWORD values that only those with the source code, and entire debugging teams, can determine for you.

You DO have to be just as intelligent as a Unix admin to setup a reliable Windows network, since the fundamental laws of computing still apply. Then, on top of all that, you have to memorize the magic Microsoft spell-book, learn all the magic incantations if you want to do anything other than the very basic default settings (eg., for when they perform badly, are terribly insecure, or just completely blow up on you, like the above).

There's really no debate about it. Samba has a config file that stays under 1K even in the complex setups. With Windows, you've got a 20MB+ registry to look through, and absolutely no way to know what each value might do... Even if you had to look through the full Samba source code, it wouldn't come close to being as cumbersome as the Windows registry, AND you'd actually know everything there is to know about it, rather than some anecdotes here and there, you get from the Microsoft spell book...

I say this as someone who knows it all extremely well... Admin for over 100 Windows systems for a lot of years, from NT4.0 to 2000, and 2003. More recently, doing everything I can to get positions as a Unix admin, even taking a pay cut, rather than put up with the nightmares that a Windows admin position brings. I still occasionally get dragged back into the Windows world, when my current employer has something that the idiot MCSEs can't figure out, on our dwindling number of Windows systems, or when a former employer or someone else who knows me by reputation gets desperate and throws enough money at me to come back and fix some show-stopper on the Windows systems I long ago told them they needed to get rid of.

In short, I agree with the GP. Windows Admin is inherently a job where you can only peek through the keyhole and turn one screw at a time, while a Unix admin can strip out entire components and work with them in isolation. A Linux (or BSD, or OpenSolaris) Admin in particular can get the blueprints to everything, and KNOW exactly how everything works, with a fairly modest amount of effort... Even if you were unbelievably lucky

Re:Computers keep getting faster

[somersault]

I think power requirements are probably the main problem, rather than the hardware. It must be pretty trivial to add more cores to a system that's already using tens of thousands of them, but you're going to need a lot of power.

These systems are only really getting "faster" for parallel tasks too - if you gave them a sequential workload then I assume they would fare worse than a high end gaming machine!

By Processor

[TheRaven64]

The view by processor is quite interesting. AMD has the top spot, but the majority of the top 500 have Intel chips. There are still two SPARC64 machines in the top 100, and a third one down at 383. All three SPARC64 machines are in Japan, which isn't entirely surprising. IBM makes a good showing, but it's interesting to see how far behind x86 they are, in a market that was traditionally owned by non-commodity hardware.

Re:By Processor

[pwilli]

I would have expected more AMD-based systems in the top-100, because super computers are usually built with cheap and moderately fast Processors, the market segment where AMD gives lots of bang for the buck.

Re:By Processor

[TheRaven64]

It's especially interesting for two reasons. Firstly, because at that sort of scale interconnect throughput and latency can make a much bigger difference than processor speed. With HyperTransport, AMD has had a huge advantage over Intel here (IBM also uses HyperTransport). It looks like QPI might have eliminated that advantage. Beyond that, you have the supporting circuitry - you don't just plug a few thousand processors into a board and have them work, you need a lot of stuff to make them talk to each other without massive overhead.

The other interesting thing is that the Chinese are using Intel processors at all. I would have expected them to use Loongson 2F chips, or Loongson 3 if they were out in time. I'm not sure if Loongson wasn't up to the job, or if they had some other reason for using a foreign-designed chip.

Re:By Processor

[Entropius]

If you're Intel you have more money to spend on marketing, which means "we'll give you a cut rate on a lot of 10000 processors just so we can have the bragging rights."

Re:By Processor

[Pharmboy]

Wikipedia shows the highest performing Loongson system before April scored 1 teraflop peak, and "and about 350 GFLOPS measured by linpack in Hefei". Sounds like they are focusing on performance/watt more than being the fastest, from a read of the rest of the article. Still pretty fast stuff, considering their newest system has 80 quads and is claimed to have a peak around 1 teraflop.

Re:By Processor

[Jeremy Erwin]

What's even more interesting is that the nVidia chips that made Nebulae so fast seem to have escaped your notice.

Re:By Processor

[stevel]

System and component vendors don't make money on these "lighthouse account" supercomputer sales. My experience, having worked in the past for a vendor that did this a lot, is that they're a money-loser. The motivation is bragging rights, though that can be fleeting. I know of several times that my employer declined to bid on a supercomputer deal as it would just be too expensive.

Typically, these systems are actually sold by system vendors (Dell, HP, IBM) and not processor vendors, though the processor vendor will support the bid. That #1 "AMD" system is actually a Cray. Software also plays a large part in success or failure.

Re:Computers keep getting faster

[TheRaven64]

These systems are only really getting "faster" for parallel tasks too - if you gave them a sequential workload then I assume they would fare worse than a high end gaming machine!

I doubt it. A good fraction of them use POWER6 processors, which are still a lot faster than any x86 chip for most sequential workloads. On top of that, they typically have a lot more I/O bandwidth. They might only be a bit faster, but it would have to be a really high-end gaming rig to be faster.

Re:A 2nd "Chinese"....

[TheRaven64]

Interestingly, the Chinese machines don't seem to be using Chinese CPUs yet. I was hoping to see at least one Loongson in the top 500.

How about a direct link...

[Jah-Wren Ryel]

How about a direct link to the actual site - or even the actual list?

LINPACK

[ProdigyPuNk]

I think this is the first benchmarking article I've read in years where the organizers actually know what their benchmark program does: http://www.top500.org/project/linpack. Refreshing to see real statistics (as good as they can make them), instead of the normal crap that is most hardware articles anymore.

I wonder what kind of score these beasts would get on 3DMark ?

Should Say "Top 500 Publicly-Acknowledged Supers"

[cshbell]

The list should more accurately be called, "Top 500 publicly-acknowledged supercomputers." You can go right on thinking that the US NSA, British MI6, and even some private industries (AT&T?) don't have vastly larger supers that are not publicly disclosed.

Food? What food?

[hcpxvi]

Of the UK entries in this list, the first few are Hector (the national supercomputing facility), ECMWF, Universities, financial institutions etc. But there are also some labelled "Food industry". I wonder what I am eating that requires a supercomputer?

Re:Food? What food?

[tivoKlr]

Maybe they're using it to determine why anyone would eat Haggis.

Re:Food? What food?

[et764]

One guy I used to work with that used to work in supercomputing claimed he did one project involving aerodynamic simulations of Pringles chips. Apparently they were originally shaped like wings, and would become airborne when traveling along high speed conveyor belts. They used a simulation to find a shape that wouldn't generate so much lift.

Re:Computers keep getting faster

[Entropius]

Parallel tasks are the whole point of using a supercomputer. The gains made in speed for sequential tasks really haven't been that great; Moore's Law for sequential tasks fell apart a while back.

Being able to parallelize a task is a prerequisite for putting it on a supercomputer.

Re:SETI@HOME has 3 million or so nodes...

[TheRaven64]

Not even remotely true. The big difference is not the bandwidth between the nodes, it's the latency. Nodes in a supercomputer can exchange data in well under a millisecond. Nodes in SETI@Home can exchange information in a few hundred milliseconds. Don't think that's important? A single 2GHz core runs 200,000,000 cycles in the time that it takes to send a message between two relatively close SETI nodes. It executes closer to 200,000 instructions in the time that it takes to exchange data between two supercomputer nodes. This means that for things that are not embarrassingly parallel problems, a pair of supercomputer nodes will be up to 100 times faster than a pair of SETI nodes with identical processors. In practice, they won't spend all of their time communicating, so they'll probably only be ten times faster. Of course, when you scale this up to more than two nodes, the delays are increased a lot on a SETI-like system, so something using a few hundred nodes can be far more than only two orders of magnitude faster on a supercomputer.

Largest Pirvately Owned Supercomputer?

[Plekto]

I was curious if any privately owned(non-corporate or government) machines made the list, and where they placed.

Re:Computers keep getting faster

[Sique]

"A good fraction" in this case means: Less than 10%. In fact, only 42 out of 500 use POWER.

actual purpose

[Iamthecheese]

In years past as many as 7 out of 10 officially listed computers were for security research. Now, contrary to the article, that's down to 2.

Jaguar -- general research (http://www.nccs.gov/computing-resources/jaguar/)
Roadrunner -- security research (http://www.lanl.gov/)
Kraken XT5 -- general research (National Institute for Computational Sciences/University of Tennessee)
Tianhe-1 -- unstated
Pleiades -- security research (nukes)

"Recently expanded to accommodate growing demand for high-performance systems able to run the most complex nuclear weapons science calculations, BGL now has a peak speed of 596 teraFLOPS. In partnership with IBM, the machine was scaled up from 65,536 to 106,496 nodes in five rows of racks; the 40,960 new nodes have double the memory of those installed in the original machine"

Intrepid -- General research
Ranger -- General research
Red Sky -- General research

It makese me wonder whether the machines for nuclear research went underground or maybe it just doesn't take a top ranking supercomputer to calculate a nuclear explosion anymore.

Re:actual purpose

[zeldor]

pleaides isnt nukes, its nasa. airplanes and weather.
the others some are nukes some are open unclassified uses.
noaa/nsf/etc

Re:actual purpose

[rdebath]

As I understand it most of the nuclear research simulations that it would be nice to run simply cannot be done on any modern machines. If it's only a few particles they can be simulated on a laptop but the interesting interactions need to simulate millions or billions of points with every single one of them influencing every other one in the simulation.

As a simple example, a genetic algorithm was used to program some reconfigurable FPGA chips. A layout was grown on the chip the did the job but broke just about every rule for FPGA design. There were parts of the layout on the chip that were not connected to any circuit but removing them made the device fail to work. Transferring the layout to a different chip got you a non-working circuit. It would be great to be able to simulate this ... not a chance it's too big, by so very many orders of magnitude.

http://www.netscrap.com/netscrap_detail.cfm?scrap_id=73

Re:SETI@HOME has 3 million or so nodes...

[compro01]

If you're gonna open it up like that, Folding@Home would almost certainly take first place.

http://fah-web.stanford.edu/cgi-bin/main.py?qtype=osstats

Re:Computers keep getting faster

[wagnerrp]

Parallel tasks are the whole point of using a supercomputer.

Well it is now. The original supercomputers were based around a single very fast processor, and had a number of co-processors whose sole purpose was to offload IO and memory prefetch, so the CPU could churn away without interruption. Modern out-of-order CPUs are effectively an old style supercomputer on a chip. Heavy use of parallel processing didn't really take off until the late 80s. This paradigm shift is what caused the supercomputer market crash in the 90s, as development devolved from custom CPUs, to throwing as many generic cores at the problem as you can and using custom interconnects to mitigate parallel overhead.

welcome to 1995

[Colin Smith]

um. you want a Beowulf with that?

Linux has been in the supercomputer lists for decades.

Google is a much better example of how you can use Linux to take over the world; which is what every self respecting middle manager want's to do.

I.e. Shit loads of cheap compute power. Got any tasks which need that?

Interesting...

[CCarrot]

"It's measured against a theoretical benchmark - if you ran a real-world application you might get a very different answer".

Next bulletin:

"Vista-based benchmark testing complete - converts Jaguar to big pussycat"

;o)