Slashdot Mirror
Efficient Supercomputing with Green Destiny
gManZboy writes: "Is it an oxymoron to have an efficient supercomputer? Wu-Chun Feng (Los Alamos National Laboratory) doesn't believe so - Green Destiny and its children are Transmeta-based supercomputers that Wu thinks are fast enough, at a fraction of the heat/energy/cost, according to ACM Queue." 240 processors running under 5.2kW (or less!) is nothing to sneeze at. The article offers up this question: might there be other metrics that might be important to supercomputing, rather than relying solely on processing speed?
I knew that sword was beefy, but that's insane!
How much of a footprint and weight they take up as a metric to consider? ;)
Join the TWIT army now!
The MHZ war has been going on for soooo long that everyone just excepted that faster MHZ related to faster machines. Well, 64Bit computers are placing chip manufactures in a position where they have to market on a platform that declares that MHZ doesn't really matter.
I think the question is a bit naive though as everyone knows a hundred software tools to rate performance of CPUs rather than just relying on MHZ.
Nick Powers
Encryption: I may not agree with what you say, but I will defend your right to encrypt it...
Why bother? If you have to sacrifice computational power for energy efficiency, then what is the point of having a supercomputer? Isn't compute power the whole purpose of having a supercomputer?
While obviously there is a bit of hyperbole in your statement (I highly doubt there are many systems defined as "supercomputers" that consume less than 3 digits or so of kW...certainly not 400-500W that a worst case enthusiast consumes), I really wonder if home computing has really gotten that much worse. Around 11 years ago I remember getting a 350W power supply for my 386-33 (with Diamond Speedstar 24x!), and this was pretty much par for the course - of course the CPU itself consumed much less power (I think around 5W) than some of the high end CPUs (that can consume up to 80W), however the board and surrounding hardware generally consumed a lot more power then.
I was talking to a friend the other day about a bunch of lab computers that my school is getting rid of - a bunch of old Pentium MMX's. He suggested turning them into a cluster. But after thinking about it, I realized that the group of about 10 old computers we had would consume more power - and would likely be considerably slower than a single one of the 2.4Ghz Dell's that replaced them. "What's the point?" I said.
Applying that here, the little VIA chips run at roughly the speed of a Celeron 500 or so, I'd say something like an AMD Athlon 3GHz would be just about as fast as about 6 of the VIA chips. So you are still saving some power, but as not as much as it would seem as first, as you need many low power chips to equal the speed of one faster chip. Not to mention power consumed by having more motherboards, network cards, switches, and other associated hardware.
Something to really look at is the cluster of G5's. The G5 chips use a lot less power than their x86 counterparts. I bet that cluster of G5's is probably right up there in terms of processing power per watt as this VIA super computer. And it's way more cool to boot.
The other point is: how expensive it is to support a cluster ? Not only the energy consumption, but also the infraestructure. It is pretty darn difficult to keep a thousand processors cold. You may need a special building, special power supply for it, etc.
A final point: as far as I know, the rule of thumb is that the floating point performance with these energy efficient processors is of the same order of magnitude as regular processor, may be a factor 2 difference.
You do the math ... :-)
Robert Cringely pointed out the benefits of this tradeoff (pure speed vs. low heat/hihg maintainability), pointing to Google's use of Pentium III-s for their server farms.
Why are supercomputers primarily benchmarked by their speed? The answer comes when you consider that almost all labour-saving devices are measured in the work they perform in a given period of time.
Time is the only truly finite resource from a human perspective. As technology has progressed, distances have been conquered, vast energies harnessed, but old Father Time is still inescapable.
As a result, we place great value on just how much time is taken to accomplish anything.
with the centaur C5P processor core. Draws about 8W for the chip @ 1Ghz. Lets assume 12W total for network boot.
p g ]
[ see image here: peertech.org/hardware/viarng/image/nano-itx-c5p.j
With 5,200 Watts for Green Destiny, you could use 433 boards these boards for the same power consumption.
The on chip AES is clocked at 12.5Gbps, Entropy at 10Mbps (whitened). Thus you would have
422Ghz of C5 processor power
5.412TB/s of AES (yes, terabytes)
4.22Gbps of true random number generation.
Yeah, these are really rough estimates, but that is a long of bang for your kilowatt buck no matter how you slice it.
With a cutting edge P4 approaching 100W the efficiency of these less powerful but fully capable systems will become increasingly attractive.
I would not be surprised to find bleeding edge processors relegated to gamers and workstations as most computing tasks start migrating towards small, silent, low power systems that simply *work* without eating up desk space, filling a room with fan noise and driving the electricity bill higher with continuous 100's of W draw.
Lottsa years ago I used to maintain a CDC 7600, not only did it need full refrigeration, but it's original design spec was for an MTBF of 15 hours! The designers reckoned that it was so fast that the biggest job imaginable could be run in that time. Of course it did better than that in the end, but it was a bugger of a job to fix, and the backplane was 6 inches deep in twisted pair wires. Just imagine making wiring changes.
The article offers up this question: might there be other metrics that might be important to supercomputing, rather than relying solely on processing speed?
Um, yes?
If you do the math with X (10,280 instead of 13,880 performance, 1000sq instead of 21,000sw, and 800kw instead of 3,000kw) you get a 337 fold increase in performance per square foot, rather than 65, and an 832 fold increase in performance per Watt, rather than 300 fold, vs the Cray.
Of course I dunno the numbers for the Transmeta solution yet!
GPL Deconstructed
Its not CPU speed that is important in supercomputer/clusters it is the speed at which you can get data from one node to esp memory access. If you havea 512 node system and node 3 needs a copy of node 40's memory it has to copy it over.
If its even just 512Mb of Gigabit ethernet and assuming 100% performace it would still take 5 seconds which is many orders of magniture. Just look at SGI machines which use NUMA and their Cray-Linux are 3.2 TeraBytes (bytes not bits). Now thats how you want to shift data
Rus
Cheap UK and US VPS
Money is usually finite, too. Especially in research. Power costs money. Cooling also costs money.
Especially when simulating nuclear weapons.
-Shane
I love teh int4rw3b!!!!!111one1
It's still valuable to have one or a few really friggin' fast processors versus a whole lot of smaller processors if you're running tasks that can't easily be subdivided. This is why people are still buying single processor PCs rather than multiprocessor boxen. If you're buying the setup for a specific purpose and multiple slower CPUs will do the job for you, then that's great; but you'll get more flexibility with speedy processors.
If one can pack the processors more densely, it would cut down on the wiring etc, or allow much shorter paths between nodes (better still, one might be able to stuff many processors on the same board or something), thereby increasing bandwidths (when you try to increase bus speed, path length and related current leakages etc do pose problems). This in turn means computations that require more 'random' communication between nodes can speed up. I suppose that's definitely worth pursuing for the more fine-grain computation where communication bandwitdh is the bottleneck.
That is the only advantage of using a Transmeta CPU. Wouldn't it be more efficient to just use a regular VLIW CPU without all the x86 code morphing stuff?
That only shows how timely the definition of a supercomputer is. 100 common desktop machines are very uncommon and obsolete 3 years from now.
I think energy efficiency (MOPS/Watt) is a very relevant metric. The reason why my PDA cannot do wideband software radio or anything that needs lots of GOPS is energy-efficiency. If the same PDA could carry 100 XScale processors instead of 1 with the same battery lifetime, I'm sure we'll have applications for it in no time.
I'd like to see an analysis that allows you to cost (i'd say price, but its not just about money) the different components of a supercomputer and account for things like power, cooling, weight, size, infrastructure etc. The factors would have to be weightable so that you can assign varying levels of importance(like if space is more precious than money). It wouldn't need to be indepth or terribly exact, but i think it would help bring out the best possible choices.
[Fuck Beta]
o0t!
Making a case for Efficient Supercomputing
From Power
Vol. 1, No. 7 - October 2003
by Wu-Chun Feng, Los Alamos National Laboratory It's time for the computing community to use alternative metrics for evaluating performance.Motivation
A supercomputer evokes images of big iron and speed; it is the Formula 1 racecar of computing. As we venture forth into the new millennium, however, I argue that efficiency, reliability, and availability will become the dominant issues by the end of this decade, not only for supercomputing, but also for computing in general.
Over the past few decades, the supercomputing industry has focused on and continues to focus on performance in terms of speed and horsepower, as evidenced by the annual Gordon Bell Awards for performance at Supercomputing (SC). Such a view is akin to deciding to purchase an automobile based primarily on its top speed and horsepower. Although this narrow view is useful in the context of achieving performance at any cost, it is not necessarily the view that one should use to purchase a vehicle. The frugal consumer might consider fuel efficiency, reliability, and acquisition cost. Translation: Buy a Honda Civic, not a Formula 1 racecar. The outdoor adventurer would likely consider off-road prowess (or off-road efficiency). Translation: Buy a Ford Explorer sport-utility vehicle, not a Formula 1 racecar. Correspondingly, I believe that the supercomputing (or more generally, computing) community ought to have alternative metrics to evaluate supercomputersspecifically metrics that relate to efficiency, reliability, and availability, such as the total cost of ownership (TCO), performance/power ratio, performance/space ratio, failure rate, and uptime.
Motivation
In 1991, a Cray C90 vector supercomputer occupied about 600 square feet (sf) and required 500 kilowatts (kW) of power. The ASCI Q supercomputer at Los Alamos National Laboratory will ultimately occupy more than 21,000 sf and require 3,000 kW. Although the performance between these two systems has increased by nearly a factor of 2,000, the performance per watt has increased only 300-fold, and the performance per square foot has increased by a paltry factor of 65. This latter number implies that supercomputers are making less efficient use of the space that they occupy, which often results in the design and construction of new machine rooms, as shown in figure 1, and in some cases, requires the construction of entirely new buildings. The primary reason for this less efficient use of space is the exponentially increasing power requirements of compute nodes, a phenomenon I refer to as Moore's law for power consumption (see figure 2)that is, the power consumption of compute nodes doubles every 18 months. This is a corollary to Moore's law, which states that the number of transistors per square inch on a processor doubles every 18 months [1]. When nodes consume and dissipate more power, they must be spaced out and aggressively cooled.
Figure 1
Without the exotic housing facilities in figure 1, traditional (inefficient) supercomputers would be so unreliable (due to overheating) that they would never be available for use by the application scientist. In fact, unpublished empirical data from two leading vendors corroborates that the failure rate of a compute node doubles with every 10-degree C (18-degree F) increase in temperature, as per Arrenhius' equation when applied to microelectronics; and temperature is proportional to power consumption.
We can then extend this argument to the more general computing community. For example, for e-businesses such as Amazon.com that use multiple compute systems to process online orders, the cost of downtime resulting from the unreliability and unavailability of computer systems can be astronomical, as shown in table 1millions of dollars per hour for brokerages an
[Fuck Beta]
o0t!
The heat output is precisely the same as the power input. All electrical power used by the PC is eventually converted into heat in the room, so a 450W PC consumes 450W of electricity and provides 450W of heat.
Incidentally, if you have a 500W heater in your room, you could replace it with a 500W PC for no extra electrical cost, and the same effect in terms of keeping you warm. Heat can be a good thing !