Slashdot Mirror

← Back to Stories (view on slashdot.org)

Can You Imagine a Beowulf Cluster of These?

Posted by michael on from the okay,-it's-a-bad-joke dept.

Hell O'World writes: "Scientific American has a story on the history of Beowulf Clusters. It's written by the guys who built the Stone SouperComputer at the Oak Ridge National Laboratory." Nice example of how old hardware can be put to use to make new breakthroughs. Nothing radically new, but hey, it's 4AM Saturday morning, what do you expect. :)

32 of 71 comments (clear)

  1. I like the incremental upgrades by Anonymous Coward · · Score: 4

    The incremental upgrades that they do are pretty slick. The StoneSouper is a heterogeneous cluster. Each hour the machines do a speed analysis on themselves to track their performance, then when new equipment comes in, they know which nodes to replace. Makes sense because at some point in time the electricity used by the slower nodes becomes very costly in relation to the amount of work contributed to the cluster as a whole. It is real slick, and as they point out, the cluster is always improving. It becomes better and better over time, unlike some conventional supercomputers which slowly lose their value over time.

  2. Re:Surprised? by sharkey · · Score: 2

    Don't you mean a 20 inch endoscopy?

    --

    --

    --
    "Outlook not so good." That magic 8-ball knows everything! I'll ask about Exchange Server next.
  3. 4 AM Saturday morning as opposed to... by Katravax · · Score: 4

    4 AM Saturday afternoon? :)

  4. Re:There's something I still don't understand by dublin · · Score: 2

    I still don't see the difference between what is described and what a well-administered UNIX computer lab on a private network should be, with PVM installed, consoles removed, and login authorized only on a single node.

    In truth, there really isn't any difference. Beowulf is the name of a concept, not a product, package, or even method. (Beowulf clusters can be based on PVM, MPI, or other message backbones, for instance - although PVM is probably the most common, there's no standard for Beowulf clusters.)

    BTW, this stuff works - I replaced a Cray with a cluster of FDDI-connected RS/6000's nearly a decade ago and today, we use Intel boxes or an IBM 390 CMOS mainframe, depending on the job to be done. (You'd never know I'm a Sun bigot from reading that last sentence...)

    --
    "The future's good and the present is nothing to sneeze at." - Roblimo's last ./ post
  5. Re:Surprised? by Old+Wolf · · Score: 2

    6AM pictures of Natalie Portman
    7AM hot grits for breakfast.

  6. The Ultimate Beowulf Project...Consciousness!!! by cybrpnk · · Score: 2

    What I would like to see happen with beowulf research is for everybody to connect their clusters via the internet (the proverbial Beowulf of Beowulfs) and go for the Holy Grail - the simulation of consciousness within the human brain. Folks, we are already using clusters of computers to search for consciousness...that's what SETI@Home is. But space is big, and very empty, and our odds of finding ET are small. Let's start a public project to search for HAL instead. Here is what we know about the brain. Here is the place to scratch the surface on how we think it generates consciousness. We are geeks, we are hackers, we joke endlessly anout Beowulfs of Beowulfs...what are we waiting for?

    1. Re:The Ultimate Beowulf Project...Consciousness!!! by cybrpnk · · Score: 2

      Ummmm...good point. 'Cept I never told anybody I thought it would be easy...and I would substitute the word "new" for "unrealistic" in your post...and I never claimed "it" (consciousness simulation) was being done on a small scale; SETI@Home is a large scale data reduction project, not a large scale simulation like I'm thinking about. Despite their fundamental differences, both projects have (would have) the ultimate goal of discovering another consciousness besides our own.

      Look, I'm not the Shell Answer Man here. If I knew how to build a positronic brain, I'd do it, start U.S. Robots of Asimov fame and be richer than Bill Gates in six months. (I wish). But neural nets, and simulations of neural nets, is a valid area of research that Beowulfs could contribute to. Do I think the hackers are all going to get togther and create HAL on the net? Nah. But discussion about what IS reasonable and IS feasible could lead to some project that would attempt to simulate maybe 10 seconds of consciousness (and not necessarily in real time, either!) for some specific situation - fight or flight, arousal, cognito ergo sum...whatever. I think it will involve a continuous (and again, not necessarily real time) input flow of "sensory data" that results in an UNEXPECTED RESPONSE from the system as a first cut as to what defines success.

      Slashdot is a forum of people for whom new ideas is the stock-in-trade. All it takes is one new idea - or one connection between the links listed and somebody who didn't know those links existed - to result in some incremental progress, however small. And the future is the summation of small increments like that.

      As for my nick, I originally wanted joto but I saved it for you and went with cybrpnk instead.

    2. Re:The Ultimate Beowulf Project...Consciousness!!! by joto · · Score: 2
      what are we waiting for?

      We are all waiting for people with nicks like cybrpnk, telling us how easy it would be to make a simulation of something nobody understands, by linking all the computers in the world together in unrealistic ways in order to achieve an ill-defined and unverifiable goal because he claims it is already being done in the small scale (without success, I might add), without explaining what would make it happen simply by adding more computers.

    3. Re:The Ultimate Beowulf Project...Consciousness!!! by big.ears · · Score: 3
      They are already doing this here, here, and here, using macs and pcs and unix boxes (UNclustered) to run "Cognitive Architectures"--simulated virtual agents that (to one extent or another) behave as real people do in simple and complex virtual environments. The problems that are being addressed out there do not require as much computing power as you might think, and the research is studying complex tasks (flying airplanes, air-traffic control, learning, memory, etc.) There is little brute-force search required in the search for 'consciousness', which is what these distributed client systems (ala SETI@home and the gene folding project) do best. The largest leaps forward have been made looking at small manageable problems that don't generally require a supercomputer. If you were able to create a giant distributed model of the brain, it very likely would be equally as difficult to understand as our own brain is; and in order to build one (for spectacle's sake or something), you would need to know a lot about the details, like local connectivity patterns.

      That being said, I don't think there's any theoretical reason someone couldn't build a fairly realistic highly-complex "brain" using, say, 100,000,000 simplified neural units (I've heard of a guy in Japan who is doing such a thing), but I don't really know what it would do, or if it would teach us anything that is interesting.

  7. The Ultimate Beowulf Project -The Shell Answer Man by cybrpnk · · Score: 2

    And for you young whippersnappers out there that don't remember the day Kennedy got shot and the rest of the 1960s (what a decade) here's the scoop on the Shell Answer Man, which would be a pretty good Beowulf project in and of itself....

  8. The Ultimate Beowulf Project...Shell Answer Man! by cybrpnk · · Score: 2

    Still more on the Shell Answer Man!!!

  9. Re:Posting an article by cperciva · · Score: 2

    Hey, slashdot moderation works!

    At least, that's the only conclusion I can draw from the fact that a /. EDITOR seems to have a karma of below 25.

    Hmm... wait a moment. Maybe moderation doesn't work. Given all the trolling, redundant stories, and flamebait produced by the editors, I don't know if they should even be posting at Score=1.

    --
    Tarsnap: Online backups for the truly paranoid
  10. Small compared to clustrum by Wizard+of+OS · · Score: 2

    This cluster is relatively small: 133 nodes. At my university they had a 250-node cluster up for a day during the lustrumcluster project. They wanted to have a 365-node cluster (the number of years that the university exists), but had trouble getting enough machines. A friend of mine (who was a member of the group that built it) told me that it equaled an 99Ghz intel machine (they used 250 intel machines with procs around 400 Mhz)

    So, this cluster is relatively small :-)

    --

    --

    --
    If code was hard to write, it should be hard to read
  11. Posting an article by Traxton1 · · Score: 3

    OK, where is it written that you must post an article at 4 AM anyway? Will CmdrTaco just beat you mercilessly if you don't? If its something you realize you have to apologize for, try just not posting it. This article I don't particularly mind, but just for future reference. Thanks.

    --
    Twitter.com/TrentonHyatt
    1. Re:Posting an article by kilgore_47 · · Score: 2

      I only watch pornography for the articles.

      I only read /. for the porno links.

      ___

      --
      ___
      The way to see by faith is to shut the eye of reason. --Ben Franklin
  12. What would you do with a cluster? Seriously. by green+pizza · · Score: 2

    What would you do with a cluster? Seriously.

    I hear this A LOT. There are even ISPs in my area that advertise their clusters. I honestly can't think of a reason why the -average- geek (and by definition, most of us geek -are- average) would need a cluster in his home, especially with PCs with CPU clockspeeds in the GHz range. Perhaps an MP3 machine in the living room and a nice fileserver tucked away in the wiring closet, but no need for a cluster.

    That's not to say clusters aren't of use.... they're wonderful cheap solutions for certain supercomputing tasks and are awesome for most forms of rendering. BUT... that doesn't mean every one of us needs an 8-node cluster to render the 3D images we make in Blender about four times a year. Nor is a cluster a magic fairy wand that can turn a pile of older, slower PCs into one magical, fast, new PC. It's not that easy. There is no free lunch. Programming for parallel processing is no simple task and adapting an existing project for a cluster can be extremely painful work at best. I'll leave "distributed GNOME" to someone else.

    1. Re:What would you do with a cluster? Seriously. by green+pizza · · Score: 2

      If I were an ISP, I'd have one for the reliability and make damn sure I advertised it.

      A high availability cluster is just that, a "high availablility cluster". I was refering to computational "Beowulf-style" clusters. There is an ISP in my area that brags about their 24 node computational cluster. I can't imagine what they could possibly be using it for, other than maybe datamining their logfiles to figure out trends in their customers' surfing habits.

      Getting back to high availability, I totally agree. Anyone with a mission critical server setup really outta have a hot spare / mirror of each machine ready to take over should the first one fail. Or better yet (if the application can handle it) use a load balanced system to allow machines to be added and removed on the fly.

      There is a *major* difference between "pure" computational and high availability clusters, though each can borrow aspects from the other. Simply setting up a beowulf-style cluster will not automatically solve every problem with availability and scaling. I don't know how many times I've seen people asking "how do I install Apache, Sendmail, and mySQL on my beowulf cluster". It's not that easy. Beowulf is not a magic fairy wand that turns a stack of PCs into one supercomputer.

  13. Re:There's something I still don't understand by green+pizza · · Score: 2

    Just what is a real cluster? I've looked through all "Beowulf" documentation I could find...

    I don't belive there is a *proper* published, accepted usage of the term, "Beowulf". And you're correct, it is interesting how the kids are just now discovering distributed computation.

    It may also be worth noting that "Beowulf", when used in a converstation, is often followed by "Schweeet", and is mostly said by l33d d00d wannabes in their prepubescent years. (Or by buzzword kings of any age).

  14. Re:Scary by green+pizza · · Score: 2

    What is more important for a render farm, the processor or the graphics card? Because when you distribute something for rendering across the nodes, would it be handled by the processors of the node or the GPU on the graphics card?

    CPU and Memory bandwidth, by far. You don't even need a graphics card on cluster nodes, as the graphics will never be used. Rendering is purely a CPU operation and thus is slow, but results in a far better image than your graphics card could ever dream of producing. The best your graphics card can do is lower-quality real-time previews. The better the graphics card, the better quality of the real-time imagery. A Quadro2 Pro, FireGL4, or Wildcat II can produce some mighty nice real-time images when modeling and previewing your 3D graphics, but are nowhere near as nice as the final rendered product. There really is no way to put your graphics card to use as a "coprocessor" in the rendering process, either, as rendering software and your graphics card differ greatly in the way they produce their images. Think of your Geforce 3 as "quick, lossy, and cheap" and a render sever (or cluster) as "slow, lossless/beautiful, and expensive".

  15. Re:Render speed by green+pizza · · Score: 2

    But what about those $1000+ graphics cards from 3dlabs? How do those help render scenes?

    See my previous comment. The expensive graphics cards only produce better quality real-time previews. They don't (and can't) assist in rendering, which is a totally differnet process that uses a different technique to handle reflections, refractions, shadows, etc.

    The same applies to even a $$$ Millions SiliconGraphics Onyx2 or Onyx3000 "Reality Monster" (1 - 16 graphics pipelines, 2 - 512 CPUs, 64 MB - 1024 GB RAM, taking up 1 - 24 *racks* of space). This beast is a single machine and can drive up to 128 independent (different view/angle) monitors or projectors (8 displays x 16 pipelines) and can handle 9 GB of gfx ram (4 GB texture ram + 5 GB of framebuffer). Yet... even this beast can only use its CPUs for photorealistic raytracing rendering. The InfiniteReality3 graphics pipes can only be used for real-time graphics (simulations, flythrus, low-quality previews of animations and movies, etc). When it comes to the final high quality rendering, only the CPUs are put to use.

  16. Scary by green+pizza · · Score: 3

    What truly scares me is how folks like this are lusting over something they could never even utilize. I could *maybe* understand someone's desire for a Cray (and even that's a stretch), given the company's interesting history (and even more interesting founder, Seymour Cray). But a cluster of PCs?? What's so great about having racks upon racks filled with x86 systems sitting in your den that will recieve little or no use? And really, what's so great about a Cray in your own home? What are you going to run on either? Distributed TTYGNOME? And no, you're not going to be able to recompile the Linux kernel in two seconds with -any- cluster. You're better off builing a nice desktop PC and a companion server. Lust over the 100 GHz PC that you'll be running in just 8 years.

  17. s/beowulfcluster/cluster by green+pizza · · Score: 4

    For the 6.022x10^23rd time, not all clusters are "Beowulf Clusters".

    1. Re:s/beowulfcluster/cluster by Daniel+Merritt · · Score: 2
      No, but had you read the article you'd discover that the clusters it talks about are beowulf clusters.

      The subtitle: ..."connect ordinary PCs so they can work together"...
      Later on: "Since them, the name [beowulf] has been widely adopted to refer to any low-cost cluster of widely available PCs".

  18. Macs also... by Johnny+Mnemonic · · Score: 2


    I didn't see any mention of the best Beowulf clutstering project of Macs, so to "Think Different": Project Appleseed, put together by the physics department at UCLA. They've accomplished phenomenal speeds, etc.; mostly, it's just as possible on Macs as it is on anything else.

    In fact, they have even developed a drag-n-drop interface for setting up Beowulf jobs.

    --

    --
    $tar -xvf .sig.tar
  19. nice homepage by Technodummy · · Score: 2

    I didn't think anyone still listened to the Beatles...

    1. Re:nice homepage by Technodummy · · Score: 2

      rum to wine is not a good switch... *L*

      Imagine is my favourite from John Lennon

  20. Re:it's 6pm Saturday you American freaks! by Technodummy · · Score: 2

    *whispers* a little known secret about beer... if you stay single, the last cold beer is always yours ;o)

  21. it's 6pm Saturday you American freaks! by Technodummy · · Score: 3

    *hehe* c'mon... catch up... we're going out drinking...

  22. Can you imagine... by Vuarnet · · Score: 2

    a Beowulf Cluster of suggestions to /. for future articles?
    - "First Posts!" has been mentioned already.
    - Natalie Portman has been mentioned already.

    So that leaves us with:
    - Hot Grits and their influence on the liberal mindset, by JonKatz
    - Microsoft sucks!, by Hemos
    - Penis birds: extinct, or merely migrating somewhere else?, submitted by the goatse.cx guy (rejected)

    Uhh... let's leave it at that, for now.
    Tongue-tied and twisted, just an earth-bound misfit, I

    --
    Tongue-tied and twisted, just an earth-bound misfit, I
    Learning to fly, Pink Floyd.
  23. There's something I still don't understand by Soft · · Score: 2
    Just what is a real cluster? I've looked through all "Beowulf" documentation I could find, and I still don't see the difference between what is described and what a well-administered UNIX computer lab on a private network should be, with PVM installed, consoles removed, and login authorized only on a single node.

    Am I missing something?

  24. God help us if computers ever unite against us by archnerd · · Score: 3

    When computers get together to make decisions, they form a Beowulf cluster. When humans get together to make decisions, they form a comittee. Need I say more?

  25. Surprised? by JBowz15 · · Score: 3

    4 A.M., Saturday - Article on the history of Beowulf clusters.

    Now, let's look into the future....
    5 A.M., Saturday - Article on the history of First Posts.

    Who know knows what 6 A.M. will hold?