Follow Slashdot blog updates by subscribing to our blog RSS feed

 



Forgot your password?
typodupeerror
×
Supercomputing Power Hardware

Creating a Low-Power Cloud With Netbook Chips 93

Al writes "Researchers from Carnegie Mellon University have created a remarkably low-power server architecture using netbook processors and flash memory cards. The server design, dubbed a 'fast array of wimpy nodes,' or FAWN, is only designed to perform simple tasks, but the CMU team say it could be perfect for large Web companies that have to retrieve large amounts of data from RAM. A set-up including 21 individual nodes draws a maximum of just 85 watts under real-world conditions. The researchers say that a FAWN cluster could offer a low-power replacement for sites that currently rely on Memcached to access data from RAM."
This discussion has been archived. No new comments can be posted.

Creating a Low-Power Cloud With Netbook Chips

Comments Filter:
  • Cloud? (Score:1, Insightful)

    by Anonymous Coward

    Didn't we have another term for this before all this cloud hype?

    Imagine a beo... Beef? Bud?

    I can't remember. My brain can't fight all these buzzwords.

    • Buzzlewolf cloister. I think. And something about running Listix.

      But no, just like every other friggin product that uses more than one CPU, it is now a 'cloud'.
  • oblig..... (Score:4, Funny)

    by omar.sahal ( 687649 ) on Thursday April 16, 2009 @06:03PM (#27604913) Homepage Journal
    imagine a beo..... oh forget it
    I tried but I couldn't resist. I reloaded three times and i was still first post
    • by Samschnooks ( 1415697 ) on Thursday April 16, 2009 @06:25PM (#27605131)
      Thank you. You saved Slashdot from this:

      I'm thinking of a Fast Array of Gigabyte Systems or "FAGS" as opposed to FAWN.

      Imagine talking to your admin in front of PC type of folks,

      "Hey Lou, you did you get those new FAGS? That last ones broke down and were a real pain in the ass!"

      "No Joe, we still have those old FAGS. The holes in those things were so big, anything could get in."

      "Yeah, I know it. They were pigs too. Some of the fuses went. Things really got blown!"

      "I tell ya! I tell ya! Hey, how are the boys in San Fransisco? I heard the FAGS vendor is really sticking it up their asses."

      "Sort of. They were happy with their shot and reciprocated on the terms."

      "Ah, good."

  • Cradle to Grave (Score:5, Insightful)

    by lymond01 ( 314120 ) on Thursday April 16, 2009 @06:13PM (#27605021)

    When I started this post, I was thinking that the overall power usage of building 21 computers that run at 85 W might supersede the power usage of building one 1000 W computer with 32 GB of memory, if you take the whole process from manufacturing to disposal.

    But I suppose it's the electric bill of the company we're concerned with so I'll just sit in the corner and re-read Bambi.

    • by maxume ( 22995 )

      If the array is cheaper to buy, probably not. Especially if it uses more materials and is still cheaper.

      I guess buying the highest performance Intel chip would throw that off quite a bit, but I doubt that is what you were talking about.

    • Re: (Score:2, Interesting)

      by drizek ( 1481461 )
      I think with the grave you are screwed either way, but with the cradle you should keep in mind that atom processors are TINY. In fact, they are one tenth the size of a nehalem processor, meaning they require about a tenth of the resources to produce. Assuming that they will be replacing a dual socket system, you break even. A quad socket system gives the atoms the win. The real problem is going to be in manufacturing all those motherboard chipsets.
      • They also don't require cooling (or very little of it compared to server CPUs), and the use of economy of scale kicks in *way* faster: how much of a price reduction would you get if you ordered them in batches of 100?

        Not to mention, the profit margins on those tiny systems inherently lower than of server hardware. Even if you used "only" 20 of them, you'd probably get more bang-for-the-buck than if you spent an equivalent amount of money on a nehalem system(s).

        The bigger issue would be networking and softwa

    • It's probably not that different. These processors have smaller dies, so making a half-dozen of them or a regular desktop CPU probably takes the same amount of power.

    • Re: (Score:3, Insightful)

      by chill ( 34294 )

      The single 1000W computer is also a single point of failure.

    • You'd have to look at computing power per watt to run, per dollar to buy and per watt of cooling, as watts are directly correlated to $$. The atoms are slow, but 21 of them at under 100W is def interesting and undoubtedly useful.

      I've been waiting for something like this, I don't think it's coincidence that Intel named this chip the atom. It's small and insignificant by itself, but add enough together and you get some interesting things....

      At what point does it become smarter to have a whole slew of th
    • by afidel ( 530433 )
      1000W, what freaking system draws 1000W and only has 32GB?!?!? One of my most power hungry systems is a DL585 G5 with 2x3.2 Ghz quad core Opteron's and 128GB of ram and a few HBA's and it draws less than half that typical.
      • Dude, SLI desktop systems can easily require a 750W power supply already. I can't see it being that hard for a beefed up system to use an extra 30%. RAM on its own isn't that power-hungry anyway, compared to CPUs.
        • by afidel ( 530433 )
          Oh yes it can be, Intel Core2 based Xeon system's fully loaded with FBDIMM's can easily use as much power for RAM as CPU.
        • Servers don't use SLI unless they're very special-purpose. Easily half the power in that sort of system is supporting the graphics card. A 1000 watt system with only 32GB of ram is either very inefficient, or doing something that a slashdot poster isn't qualified to pontificate about.

      • 1000W, what freaking system draws 1000W and only has 32GB?!?!?

        My Desktop PC and its PSU [overclock3d.net], you insensitive clod!

    • I know noone reads TFA, but it's worse when someone skims the summary but doesn't pay attention to it.

      You should have read the following sentence: "A set-up including 21 individual nodes draws a maximum of just 85 watts under real-world conditions." ratehr than skimming through it.
      • Hm, ok, I may have misunderstood you point, but only because it doens't make much sense. What matters here is whether the price difference would be compensated for from power bills over their lifetime.
    • Simple economics (Score:3, Interesting)

      by dj245 ( 732906 )
      This argument is misinformed.

      Businesses are in business to make money and put food on the table. Nobody does anything for free. If I build a widget and it costs me $10 in electricity, $5 in heating, and $3 in cooling, my widget is going to be $18 more expensive as a result. Now, I don't do things for free, so I'll just add $18 to the cost of my widget. Probably $20 because I want some more markup for my trouble.

      Energy costs are always included in anything you buy. If the initial+electrical cost of
    • its 85 watts for the entire setup.... Not per node. ;)

  • Next Generation (Score:4, Interesting)

    by glitch23 ( 557124 ) on Thursday April 16, 2009 @06:15PM (#27605043)
    This is the next generation of a Beowulf cluster using the next-generation of hardware which is cooler, cheaper CPUs and solid-state storage and memory. Someone was bound to come up with this idea because it only makes sense. It is good to know that we have a proof of concept now so someone else can take the idea and modify it to come up with something even better. Eventually hardware manufacturers will take notice and release it as COTS hardware. For companies who want cooler and cheaper server hardware this would be a good fit once it has been packaged as a COTS product.
    • Google already uses COTS hardware for their servers [slashdot.org].

      This may be a step forward in terms of modularity and scalability, however. Rather than 1AAA shipping containers, the server "batches" could be the size of refrigerators and powered by one power supply "per-fridge".

      At any rate, I especially agree with your latter statement: hardware manufacturers will be forced to take note, as this gradually becomes more common.
    • by dbIII ( 701233 )
      There was "green destiny" in 2002 running on the Transmeta Crusoe TM5600. It makes sense for the right job and these sort of machines don't need a lot of cooling.
    • by joib ( 70841 )

      Something like the SGI Molecule [gizmodo.com], perhaps.

      Oh yeah, RIP SGI.

  • AMD Geode? (Score:3, Insightful)

    by TheRaven64 ( 641858 ) on Thursday April 16, 2009 @06:19PM (#27605073) Journal
    Weird choice. I suggested a while ago using a set of OMAP3 chips in blade servers. You can get a 600MHz ARM CPU, 512MB of flash and 256MB of RAM in a single package-on-package module, with a power usage of under 1W. Put an 8x8 grid of them on a board and you've got a nice little wedge of server power at well under 64W. Use a bit SAN elsewhere in the rack and you've got a set of machines you can bring online easily for individual users. You could assign individual ones to different users / customers and just plug in more when they were needed. If I were doing it now, I'd be tempted to use one of the newer Freescale ARM designs that goes to around 1GHz and has on-die Ethernet controllers.
    • I remember there was a company that tried this a few years ago. They created a server with something like 3500 CPUs, consuming roughly 1500 watts.

      I don't believe it ever caught on. Since it wasn't x86 or ARM, porting software would probably be incredibly expensive. Also, splitting tasks between that many cores or CPU is... difficult.

      • Re: (Score:3, Informative)

        Sounds like you're referring to SiCortex, who uses MIPS chips. They also have a desktop workstation with 72 processors and 48 or 96GB of RAM that only consumes 300W. http://www.sicortex.com/ [sicortex.com]
        • Yep! I recognize the pics on their site. Was definitely them.

          Looks like they've updated their hardware - 64bit now, with close to 6k processors. (cores, probably)

  • New buzz words? (Score:5, Insightful)

    by LoRdTAW ( 99712 ) on Thursday April 16, 2009 @06:20PM (#27605087)

    So I guess the word cloud has replaced cluster to give old technology a fresh new look. Gotta love marketing.

    And since when did the term netbook come to describe low power computing hardware? There have been mini-ITX boards with low power CPU's long before the term netbook was in use. Just more marketing bullshit, repackage existing tech with a shiny new name and sell it.

    • Re: (Score:1, Interesting)

      by Anonymous Coward

      So I guess the word cloud has replaced cluster to give old technology a fresh new look.

      A cluster is a cloud when it is sufficiently large and the nodes are sufficiently small, like the water vapour of a cloud. Isn't is poetic?

      • So the massively parallel touting of this new buzzword should best be described as a 'cloudfuck'?
    • There have been mini-ITX boards with low power CPU's long before the term netbook was in use.

      Allow me to extend the above:

      There have been other boards with lower power CPUs longs before anyone cared about VIA or their mini-ITX form factor.

      • by LoRdTAW ( 99712 )

        Good point but they were targeted at embedded systems and other non-pc oriented systems. I was talking about commodity low power PC hardware.

    • So I guess the word cloud has replaced cluster to give old technology a fresh new look. Gotta love marketing. So I guess the word cloud has replaced cluster to give old technology a fresh new look. Gotta love marketing. So I guess the word cloud has replaced cluster to give old technology a fresh new look. Gotta love marketing. So I guess the word cloud has replaced cluster to give old technology a fresh new look. Gotta love marketing. So I guess the word cloud has replaced cluster to give

  • by seifried ( 12921 ) on Thursday April 16, 2009 @06:27PM (#27605145) Homepage

    Or to put it simply: pulling a "finished" object from memcached will almost always be faster then having a machine create/render/whatever you do to create the object. If you want to pull large amounts of data from RAM buy a 1U server that takes 64 gigabytes of ram for $5000 (so about $78 per gig of ram, and much faster than a compact flash card in a super cheap laptop). Or buy solid state disks/PCIe RAM cards. Now if we're talking about building a render farm for whatever (frames, objects in database, etc.) simply run the numbers, how many objects/sec/dollar do you get with different solutions and how important is latency.

    What interests me is the ease of building a many node cluster and learning how to administer and write software for something with 20+ nodes.

    Of course you could just buy computer time from amazon.com EC2 for $0.10 per hour per node and practice there ($2 an hour for 20 systems running. not bad).

    • What interests me is the ease of building a many node cluster and learning how to administer and write software for something with 20+ nodes.

      This is a chicken-and-egg question. If this becomes more common, the software (and hardware) to manage and administer these will be made available due to commercial needs, and interest from the OSS community. It's just like threading is making its way into almost every software "domain", as multi-core CPUs are becoming the norm.

      • by seifried ( 12921 )
        Agreed, but it's still nice to be able to practice, and for that you need multiple machines (e.g. variable network latency vs.s. VMware running many images with no jitter in communications, etc. transient failure conditions, you name it.).
    • Or to put it simply: pulling a "finished" object from memcached will almost always be faster then having a machine create/render/whatever you do to create the object.

      I don't think the idea is to dump the concept of cache. The idea is to drop the added complexity and expense of "memcached". Instead of retrieving data from slow power-hungry hard drives, processing it and caching it in very expensive SDRAM you employ more traditional filesystem-based caching to much cheaper flash drives. That would still be less resource-intensive than re-rendering data, even though it is slower than memcached on power-hungry systems.

      If you want to pull large amounts of data from RAM buy a 1U server that takes 64 gigabytes of ram for $5000 (so about $78 per gig of ram, and much faster than a compact flash card in a super cheap laptop).

      More importantly than "is it fast" is would it be "fa

  • Wait, what? (Score:4, Interesting)

    by Enry ( 630 ) <.enry. .at. .wayga.net.> on Thursday April 16, 2009 @06:42PM (#27605249) Journal

    256M per node times 21 nodes equals 5GB. 84 watts is nice, but I just built a home server with 4GB of RAM and 2 1TB drives that has a low power AMD chipset in it. At idle, it's about 70 watts, and gets to about 100 watts when under load. Replacing the two 1TB drives with an 80GB SSD would probably be closer to what if represented with FAWN.

    Figuring $100 for the motherboard and parts makes that total system cost $2100. My server was about $500.

    Don't get me wrong, this is an interesting idea. Using an Atom can get you a lot more performance for not much more power use, and you can go up to at least 2GB RAM per node. But there's a limit to how small you can make a single item in a cluster before you're duplicating effort without much benefit.

    • by seifried ( 12921 )
      You forgot the 4gig compact flash card in each machine.
      • by Enry ( 630 )

        No, I mentioned the 1TB drive that could be replaced with a single 80GB SSD.

    • Hell I've got an Intel Board running an e6300 (65 watt cpu) with 8GB of RAM plus a Geforce 7300GT that only draws 120 watts at full load (F@H) on the cpu. Total Cost is just under $800 with the recent RAM upgrade from 4GB to 8GB

    • Re: (Score:3, Interesting)

      by CAIMLAS ( 41445 )

      That may (and really is) true. But how well does your machine work with concurrency? Or, for that matter, how fast is the processor?

      8Gb of RAM is nice and all, especially with modern software and emulated environments. But how many

      For a web-facing system - or anything serving multiple requests per second from different locations, with multiple threads all needing a quick response - having 21 500MHz cores would be much better than having 4 2.6GHz cores. That is, provided you could handle distributing the req

      • by Enry ( 630 )

        It's a 2.5Ghz 64-bit processor, dual GigE. If you're talking concurrency, don't forget OS overhead for each of those 21 systems. Each has a kernel, cache, kernel- and user-land processes running. For modern PCs it's not that big a deal, but it adds up for tiny PCs (consider the NSLU2 which can get overwhelmed running dhcpd and bind).

        Again, I'm not saying that this is a bad idea, I think this implementation isn't much compared to what's out there now.

        • by CAIMLAS ( 41445 )

          The NSLU2 isn't exactly a fair assessment, particularly with bind. The NSLU2 is, at best, a 266MHz Xscale, which Linksys shipped underclocked to 133Mhz. Also, bind isn't exactly a light system - on my 700Mhz Celeron system, serving a small 6-host, 3-user max LAN makes bind be the highest CPU-utilizing process (often). That system also runs apache, mysql, and a small drupal install. Statistics on CPU utilization still shows bind utilizing a lot of CPU utilization.

    • Not to mention the admin complication of having to set up 21 machines rather than 1. Also, you now pretty much tie up a 24 port switch rather than 1 port on a switch.
    • But there's a limit to how small you can make a single item in a cluster before you're duplicating effort without much benefit.

      The thing is, there is vast room for improvement in the cluster concept with more current technology. If you used an ARMv7-based node you'd have better capabilities in each node than the Geode, at about $100 per node (making it cost the same as your suggestion) and a PEAK power consumption of around 30 or 35 watts (your system peaks at 100 watts, still significantly higher than the 85 watt peak of the Geode-based cluster).

      Also, slapping an SSD into a regular PC doesn't make it even close to comparable apar

    • by ravyne ( 858869 )
      The problem this addresses is bandwidth, though.

      Yes your server has dual-channel DDR2 at ... 400Mhz is common (PC2-6400), with a bandwidth of 12.8 GB/s. You've got maybe 2 Gb/s network connections, for a quarter GB/s network bandwidth, and maybe 3 Ghz of processing power to push it out.

      This array, lets assume DDR2 at 333Mhz, has a bandwidth of 5333MB/s x 21, or 112 GB/s to RAM, 1 Gb/s x 21, or just about 3 GB/s to the network, and 500Mhz x 21, or over 10Ghz of processing power to push it out.
  • I've been toying around with a Samsung 16GB SSD. Performance improvement over spinning disks in an I/O-heavy scenario was neglegible. Also, it seemed as if the Linux kernel was still using memory to buffer SSD disk I/O. Which somewhat negates the argument of using SSDs to free main memory for other stuff.

    Any idea what type of OS/filesystem combination they were using?

    • by CAIMLAS ( 41445 )

      Which somewhat negates the argument of using SSDs to free main memory for other stuff.

      You misunderstand how that's supposed to work. You don't "free main memory" to SSD. The idea is to use SSD as a pre-buffer for RAM, so it's quicker to access than reading from disk.

      You buffer from a 500Tb SAN to a 100Gb SSD, to 32Gb of RAM, to 4Mb of L3, to 2Mb of L2, to 512Mb of L1 - or whatever. You don't buffer -to- a slower device, but for a faster one, so the data will be available for the pipeline when it's needed. You want to use as much of the faster memory as possible, to increase system speed.

      Usin

      • by aharth ( 412459 )

        You misunderstand how that's supposed to work. You don't "free main memory" to SSD. The idea is to use SSD as a pre-buffer for RAM, so it's quicker to access than reading from disk.

        Sure.

        But there's something wrong if the Linux kernel buffers SSD I/O in main memory and swaps code fragments to disk. At least that's what happened in my experiments.

  • Obviously it can perform fast, but it isn't going to last too long. Maybe flash is cheap enough that its limited read/write cycles aren't a serious issue, but this thing is going to chew up flash like no one's business.

    I do like that my school's eldest beowulf cluster is now completely obsolete though, costing as much power as a few space heaters and processing as much data as a cluster of iPhones.

  • 2 buzzwords in 1 title, can we do better ?

  • Intel X25-E, 2.6 watts, 3300 Write IOPS, 35000 read IOPS*. So only one or two orders of magnitude more efficient...

    And though no prices are given in the article for the FAWN, at $800 for the X25-E it's probably less expensive too. Particularly if you include setup and administration costs.

    Not a bad idea in general, and not a bad idea in specific for 5 years ago, but pathetically outclassed in every area by a high end modern SSD.

    * http://download.intel.com/design/flash/nand/extreme/extreme-sata-ssd-datasheet [intel.com]

    • by CAIMLAS ( 41445 )

      Somehow, I think you misunderstood what this article is about.

      It's not about SSD. It's about a "cloud" cluster, performing the same amount of work as (say) a dual quad core server due to its ability to distribute load over many more cores.

      • by Makoss ( 660100 )

        Somehow, I think you misunderstood what this article is about.

        Given the very frequent mention of 'disk based storage', and how flash is so much better, I'm not sure that I did.

        It's not about SSD.

        No it's not about SSD, that is the problem, it reads like they have never heard of them.

        Memcached prevents Facebook's disk-based databases from being overwhelmed by a fire hose of millions of simultaneous requests for small chunks of information.

        flash memory has much faster random access than disk-based storage

        Each FAWN node performs 364 queries per second per watt, which is a hundred times better than can be accomplished by a traditional disk-based system

        Swanson's goal is to exploit the unique qualities of flash memory to handle problems that are currently impossible to address with anything other than the most powerful and expensive supercomputers on earth

        Swanson's own high-performance, flash-memory-based server, called Gordon, which currently exists only as a simulation...

        I'm not saying that a wide array of low-power nodes is a bad idea. But unless they address the current state of technology, rather than a conveniently quaint world in which using flash as your primary storage makes you some sort of innovator, it's hard to take them seriously.

        "you could very easily run a small website on one of these servers, and it would draw 10 watts," says Andersen--a tenth of what a typical Web server draws.

        And how does that per-website ene

  • I recently had my cat push a laptop off a desk, and break just the screen. I hooked it up to a crt, installed Gentoo, and now use it as a personal server. It's quieter, and uses less electricity. It has been a great solution for me, and wonder how many others could benefit from this thinking. But this article is sort of taking my idea a few steps further. It could work well.
    • Me too

      I used an old thinkpad T20 (P3 700Mhz, 512 RAM) and now rocking a T41 (P1.6M with 512RAM).

      Runs a full LAMP stack, squid + privoxy caching for my LAN + a left4dead server 24/7 no worries

    • Way ahead of you. I've been using old laptops as servers for years. They're small, quiet, have their own efficient UPS, and are very easy to stow in a corner and add to an established wireless network with OpenVPN. Most old ones will run on under 10W mains, which is less than many devices draw while turned "off".

    • Re: (Score:3, Funny)

      by _Stryker ( 15742 )

      Why did you have the cat push it off the desk? Were you too lazy to do it yourself?

    • Re: (Score:1, Insightful)

      Laptops make great low power servers. I've been using them for ages. No noise, small footprint and built in UPS. I'm suprised someone hasn't taken the technology and used it in datacentres (without the LCD's of course). I can easily imagine someone like google with racks and racks of $99 laptops without screens being used as nodes
      • without the LCD

        Now that you mention it, Asus actually make an extremely small PC called the 'Eee Box' which is just that; the typical netbook hardware in a tiny box with no screen. It's small enough to attach it to the VESA mount on the back of a monitor.

        • by b0bby ( 201198 )

          The problem with the eee boxes is that they cost the same as the netbook versions, but you don't get the screen... I guess it's because it's a smaller market.

          • Yeah, I wish they were a little cheaper. On the plus side, the 80GB versions are 70 euros cheaper than the 160GB ones - 280 vs 350. That puts it into the acceptability range if HDD space isn't important, I think.

            That said, even at retail a 160GB 2.5" disc is 45 euros, so you can save a little money there. Apparently it's easy to install as well, there's a hatch on the bottom for it.
  • http://gizmodo.com/5091473/sgi-molecule-packs-10000-atom-cores-one-ton-of-awesomeness

HELP!!!! I'm being held prisoner in /usr/games/lib!

Working...