Transmeta Meets Blades

The Griller writes "Gordon Bell, one of the creators of VAX, and Linus Torvalds were at the launch of a new supercomputing platform at the Los Alamos National Laboratory. Based on Crusoe processors from Transmeta and running a version of linux, it is aimed at being cheaper than conventional supercomputers by requiring no cooling and lower maintenance. " Basically, it's blade clustering, using Beowulf.

Comments

Slashdotted ALREADY?!?

[Anonymous Coward]

Oh well, here's a list of mirrors...

I've got to wonder

[Arker]

I've got to wonder why they are using Crusoes. It's a good chip for the application, don't get me wrong... but the last I heard the main advantage it has over StrongARM is x86 compatibility, which shouldn't be an issue here.

Re:I've got to wonder

[The Monster]

I've got to wonder why they are using Crusoes

Because if they used Intel chips, Transmeta wouldn't make very much money off it.&lt/joke&gt

But this does explain why it's been very important for Linus to push MP in the kernel.

Re:I've got to wonder

[gmack]

Uhh no.. this has nothing at all to do with MP since i's a beowulf cluster and last I checked you can't to MP with transmeta.

Re:I've got to wonder

[mocm]

Strongarm has no FPU and is not as fast as a Crusoe.
A Pentium III still needs way more energy than a Crusoe. You have to keep in mind that those energy savings of Intel ships are usually accomplished by lowering the processor clock rate which will not help very much if you need processing power. The Crusoe also changes the clock rate, but does so dynamically, so that you always have the speed you need. Additionally, it has far fewer transistors and therefore needs less energy even at full speed.

Re:I've got to wonder

[Arker]

According to the story, they are using the Crusoe chips because they don't require active cooling, unlike Intel or AMD chips.

Obviously you are not familiar with the ARM family of processors - they are very similar to the Crusoes, and in particular they don't require any active cooling either.

Forget Beowulf

[guttentag]

Imagine if these weren't clustered...

Cube of Crusoes

[geoffsmith]

Given that you don't need to actively cool these chips, I think what would be even cooler(N.P.I.) is a cube of chips stuck together and interwoven with some sort of vascularized heat-sink. A meaty cluster of 100 chips you can hold in your hand, and plug into a big cube-shaped socket on your supercomputing motherboard. Now *that* would be New for Nerds.

Websurfing done right! StumbleUpon

transmeta.com

[jbrw]

transmeta.com has more information on why a Crusoe based solution was selected.

It all comes down to "power consumption, size, reliability and ease of administration", apparently.

And the marketing people at RLX Technologies should be shot for not having a press release up for this, as it's all based on their product...

Re:transmeta.com

[Arker]

Anyway, I like the point about "stop using more transisters to make it go faster" bit. What a hoot. That't exactly the point of building a cluster. More chips, more transistors, more FLOPS.

That's exactly why he likes designs that don't use more transistors per cpu. The heat and power consumption of a P3-P4 class chip may not seem all that bad when you have one in your PC, but when you have 100s of them racked up it can become a very serious problem.

Green Destiny

[guttentag]

...the unveiling of the supercomputer, a Beowolf cluster called Green Destiny...

Computing legends Bell and Torvalds looked on in envy as the Green Destiny blade cluster was unveiled, knowing only the great Li Mu Bai was worthy of wielding the blade cluster's power. Upon plugging the Green Destiny in, they were appalled to find it had been "r00t3d" by some Chinese hacker calling himself Yu Jen...

Post-X86 clustering

[baka_boy]

Personally, I'd much rather have a rack of XServe 1U boxes than Transmeta chips -- G4 processors may not be quite as power-efficent as Transmetas, but they also run at higher clock speeds, have two processors per mobo, give you fast 128-bit vector processing unit (very nice for scientific calculation), and still beat the pants off of PIII/IV and Athlon chips in the power/heat/size arena.

The only trick would be getting the things to work properly in a headless configuration -- Apple won't ship them without a graphics card, but I'm relatively certain that you could get a LinuxPPC installation to work even without the card installed.

Re:Post-X86 clustering

[Hercynium]

Actually, look a little more closely at the tech specs on apple's site... it says that os X server was specially tweaked to run headless on these. (it also mentions the db9 serial port set up for the old-skool unix geeks! Yay!)

Re:Post-X86 clustering

[angel'o'sphere]

Actually, vector processing is essentially useless for most scientists as long as the compiler doesn't autovectorize the code.

Thats wrong. The rest of your post also.

double x[veclen]; // init it somehow
double y[veclen]; // init it somehow

double scalar_product = 0;

for (int i; i less_than veclen; i++) {
scalar_product += x[i] * y[i];
}

This above is scalar code. Any compiler aware of a vector processor compiles that to a singel vector processor instruction. At least that was the case 14 years ago when I worked on vector processors.

I'm not sure if Altivect is a true vector processor, I think it supports like MMX only very limited SIMD processing, but I'm not sure as I say.

Operations on "arrays", hence vector processors, are very easy to map on vector processing units.

Regardless if it is as easy as above or if you have offsets or gaps like i+=3 in the loop above.

Same is true if the result is a vector again of course.

Manual vector processing instructions get interesting if the loop aove would calculate a vector and that vector was nput for a further stage.

Like this:

Vector a, b, c, d, e;
Scalar i, j, k;

a = i*b + j*c; // result is a vector
e = a + k*d;

Ususlay you would have loops calculating that, the second loop would run after a is completely calculated.

If there is a second vector processor (or just a unit on the processor) you can feed a dirctly into it tocalculate e.

AND THIS is hard to figure for a compiler. Probably youment that. As all vector units are different in that respect there exist fortran libraries with standard subroutines for that.

angel'o'sphere

Mainframe?

[GigsVT]

How can you even compare this to a mainframe?

Clustering is a very good and very cheap and superior alternative in some cases.

In the cases where you really need a mainframe, no cluster is going to help you. Mainframes aren't even really that fast. What they are good at is having tons of I/O bandwidth, even between nodes.

If we quit comparing clusters to mainframes, then people might take clustering more seriously. They are not intended for the same classes of problems.

I have an OpenMosix cluster at home, and I work with an Origin 2000 at work. (If anyone else uses IRIX you know that you work *with* IRIX, not on it, it has a mind of it's own :) They are vastly different concepts, apples and oranges.

Interesting Wording...

[evilviper]

"Gordon Bell, one of the creators of VAX, and Linus Torvalds"

Wow, so it is true... Linus is a robot.

Re:Sorry, but ARM == No floating point hardware!

[Arker]

Ok, finally that's a legitimate response. It's true ARM doesn't include an FPU. However, the last I checked (and I'm not real up to date on it) using libfloat it had emulation good enough to keep up with IA32 fairly well on FP.

I imagine, though, this is probably the reason. It seems reasonable that Supercomputer work would require some FP, although I don't know for sure.

Why limit yourself to x86

[eagl]

Why limit yourself to the x86 instruction set when the transmeta processor just needs a new instruction set decoder to emulate pretty much ANY processor? It seems like while they'll be able to use lots of existing software out there, they could get even more performance, efficiency, or maybe just easier programming by using whatever instruction set makes sense for the project.

It's all in the pre-processing with the crusoe, x86 is just there for slideways compatibility and doesn't need to be a limiting factor. When you're using a custom computer, whether it's one or a thousand crusoe processors, wouldn't it make sense to try for some compiler efficiency based on the actual hardware instead of the 8086 legacy?

Why Crusoe? Administration Costs?

[chill]

Using this site as an example to estimate power usage, we get:
240 computer blades in Green Destiny x 6,480 hours uptime (9 months) = 1,555,200 computer hours of uptime

Assuming the only thing changed on the blade is the CPU -- and North Bridge chipset, since the Crusoe includes
a North Bridge on die and the P-III does not -- at full blast the Crusoe consumes about 1.75W of power and the
P-III + NB consumes between 4.5 - 8 W, depending on chip model. However, the 4.5W number is an approximation
from the 0.13 micron ULV P-IIIM chip running in "Battery Saving" mode, or SpeedStepped down to 300 MHz. Running
at full 700 MHz tilt, with NB, we are still talking 5.75W of power consumed.

1,555,200 * 0.0175Kw * 0.10 (dollar per KwH power cost) = $2,721.60 electricity cost/year (Crusoe)
1,555,200 * 0.0575Kw * 0.10 (dollar per KwH power cost) = $8,942.40 electricity cost/year (Intel)

A saving of approx. $6,200/year in direct electric costs.

However, the big savings comes from the heat dissipation of the units. While the newer LV/ULV P-IIIs do not require
active cooling, they still run quite a bit warmer than the Crusoe units. As a result, you don't stick a rack
full of them in a room that isn't temperature controlled. The difference in the air conditioning bill can
easily reach tens of thousands of dollars.

In business, there are two types of money/budgets. One-time grants and acquisition budgets are large chunks of
cash. Recurring expense and operations budgets are smaller. Being able to get a large chunk of cash to BUY a
cluster/supercomputer is one thing. Being able to go back year-after-year and get the funds to keep it running
is another project altogether. $15,000 - $20,000/year for electricity used in running/cooling computers is a
LOT of money to some people. This doesn't include construction or maintenance costs on a custom facility/room.

As far as reduced administration costs go, many conventional supercomputers required chilled water and other
special considerations for operation. People with experience managing things like Sun E15000s and Cray T3Es
are few and far between. They are the last of the "high priesthood" of computer administrators and cost a LOT
of money to employ.

A blade server, on the other hand, is a bunch of x86 computers running Linux -- nothing a couple of grad students
can't learn the ins-and-outs of over a term. Maintenance contracts, spare parts, etc. are also TONS cheaper for
the blade/cluster solution as opposed to high-end SGIs, Suns, Fujitsu and Cray super-computers.

Another site with a bit of good supporting information is
PC Stats.

Admin a CPU by hot-swapping it

[yerricde]

could someone explain how a microprocessor is administered?

In a large cluster, the question is not whether a processor has failed, but how many have failed. Such clusters generally make it possible to swap out a failed processor while the program is running. Chips that last longer will reduce the dependency on expensive technicians to keep coming in and swapping in new boards.

Get a clue

[Arker]

Low power consumption. A 90% power savings is sort of irrelevant with a single CPU, but talk about saving 90% of electricity across 20 CPUs, and that's a decent savings.

Get a clue. The Crusoe consumes about 2 watts. Very nice compared to Pentium-class room heaters, yes, but I asked why they choice Crusoe over StrongARM, not Crusoe over IA-32. A 600mhz SA uses 450mW, so you can run roughly 4 of them for the same power and heat as one Crusoe.

The advantages that Crusoe has are two - first, as I mentioned originally - x86 compatibility. This is not a help for a supercomputer - you're going to be compiling everything from source anyway. The other advantage, that I forgot, is that the SA doesn't have an FPU. That, at least, is a legitimate reason to consider the Crusoe, but I'm still not sure the decision actually makes sense - the SA is a very nice chip and if programmed right it should have no problem keeping up with the Crusoe even on FP, figuring that you can use 4 times as many SAs for the same heat and power requirements.

Re:Get a clue

[buffy]

Here's another advantage: server blades using the TMTA processor are already a readily available commodity device. Me thinks a big part of the paper(s) was the fact that these were off the shelves devices that they used to build the Green Destiny.

Others make Intel server blades, but I don't think I've seen any that are based on ARM.

I think that goes a lot of the way towards answering your original question.

ObDisclaimer: I work for RLX.

Alpha's might be cheaper in the long run

[Billly Gates]

I had a beowulf supercomputer designer at my linux users group and he mentioned that alpha's were the cheapest per operation to run over any other platform. This was 2 years ago so this might be a little outdated and would be cheaper today to implement but anyway he was processor agnostic but if he did the math. The processor is only a small fraction of the total cost of the system. In this guy's example for weather forcasting modeling he had a 1 to 2 gigs of ram in each node and some expensive fiber based networking cards and switches. If you do not have at least a 1gb/sec transfer rate you have a major bottleneck. Anyway an intel based solution for his 35 mode cluster added with the networking, ram and switches averaged $2,000 a node. An alpha would average close to $3,000 a node. But he would recieve close to a %50 performance gain for using alpha's. So thats a %50 gain for a %30 price increase. Sure cooling might cost more but thats tiny compared to the amount saved by the cluster finish faster.

New metrics?

[morcheeba]

> Feng also proposed that a new technique is needed for measuring the performance of supercomputers. Instead of looking primarily at how many calculations a system can run in a given amount of time, researchers should also consider factors such as downtime, size, price and maintenance requirements, he said.

Following Feng's lead, the whole supercomputing industry has reacted to this new paradigm shift. Industry leader Cray has ceased development of its upcoming SV2 and has designed a system based on the reliable commodore 64. Explained lead scientist Joel Grey, "We managed to get a C64 computer out of the dump, and bought 1,000 surplus 'Barney' solar calculators off of ebay for $30".

The new system, dubbed the SV64, is not quite as fast as the SV2, but exceeds at new metrics: Converted to run on solar power, and having spent the last 15 years in an uncooled closet continously generating the "experiencing technical dificulties" logo for a local community access TV station, the new computer shatters existing power and reliability records. "With an expected retail price of less than $1M USD, we expect this computer to eclipse [Japanese rival] NEC's lead and become the platform that will be used to perform most of the world's weather, biological, and nuclear simulations well into the next decade", said Grey.

Wall Street analysts pointed out the the system has never needed maintence, nor suffered downtime, nor needed the services of an UNIX system administrater, and as a result, the total cost of ownership should remain low. Shares of component manufacturer Commodore rose 10 points to 10 1/64 in heavy trading today.

Re:New metrics?

[morcheeba]

In other news...

Olympic Speedskating will be judged, not on speed, but on fashion, sweating the least, and the contestant who books the best airfare to the event.

North Korea became the second country to land a person on the moon and return them safely to earth. Although technically the rocket blew up on the launch pad, it was still considered a sucessful mission given the impoverished country's lack of funds, the technical embargoes placed on the country by space-faring nations, and the total lack of a Korean space program.

Life insurance companies will now pay benefits for near-death experiences, close calls, and "getting really scared".

sorry... I just prefer the normal metrics of FLOPS, MIPS, bandwidth and topology for my supercomputers...