Become a fan of Slashdot on Facebook


Forgot your password?
Google Power Hardware

Google Running 900,000 Servers 127

1sockchuck writes "How many servers is Google using? The company won't say, but a new report places the number at about 900,000. The estimate is based on data Google shared with researcher Jonathan Koomey, for a new report on data center power use. The data updates a 2007 report to Congress, and includes a surprise: data centers are using less energy than projected, largely due to the impact of the recession (buying fewer servers) and virtualization."
This discussion has been archived. No new comments can be posted.

Google Running 900,000 Servers

Comments Filter:
  • by Enry ( 630 ) <enry@way[ ]net ['ga.' in gap]> on Monday August 01, 2011 @11:05AM (#36947686) Journal

    We've moved from 1U systems with 90-125W systems to blade enclosures with 60W CPUs and also getting 4 or 6 cores per physical CPU rather than 1 or 2. While our HPC cluster core count has increased by a factor of 4 (allowing researchers to do more work), the amount of energy and floor space required did not increase that much at all.

    • by Skapare ( 16644 )
      ARM for even lower wattage.
      • by alen ( 225700 )

        and ARM is gimped in that it doesn't support a lot of features x86/x64 does that makes it run so fast compared to ARM

        biggest advantage of ARM is that the SoC includes the GPU and the RAM. this is going away. x64 is now shipping with GPU on board for most CPUs.

        • uh, I believe we're talking about server situations, not consumers. Having a GPU on an ARM chip or on an X64/x86 chip is a nonsequitur. Or did I miss something here? I fail to see where you come up with this shit considering even Intel is trying to make an ARM chip. []. You think they're doing it because supposedly arm doesn't run as well or about having GPU's on the chip? Hint: Intel is shitting their pants over ARM right now.

          • by tibit ( 1762298 )

            They should only worry about enterprise/server market when there's a full featured JVM for ARM. So far, there isn't one.

            • Curious: in what way is compiling the OpenJDK for ARM not full featured? Not supported by Oracle? Or are you referring to hardware support / Jazelle?

              • by tibit ( 1762298 )

                Yeah. Oracle supports x86 and 64 bit SPARC. The only "news" I could find about OpenJDK for ARM is that support is due to be dropped from IcedTea []. So all I know is that whatever "Java" exists for ARM is abandoned at this point. Anyone who knows otherwise, please chime in. Do note that a full featured Java should IMHO have both interpreter and JIT, and perhaps be in somewhat widespread use so that there'll be enough real-life test coverage. I wouldn't use it for any major ARM-based project at this point, unle

          • GPUs provide substantially faster floating point processing than a general purpose CPU. Putting these new Intel/AMD integrated chips in big iron supercomputers will give research teams orders of magnitude more computing power (for less power and money) than the current CPU-only based offerings.

            As far as Intel trying to make an ARM chip, that's for an entirely different market.

            • putting an integrated graphics chip onto a server cpu is a joke. Even if you put 32 integrated graphics cards onto a single 8 core server cpu, the flops will be shit in comparison to any discreet graphics card which costs orders of magnitude less.

              • I should have explained myself better. You're thinking about small blade servers doing simple tasks. Sure, for that the GPU is a complete waste. But that's not what I'm talking about. I'm not talking about a computer that's cranking out killer graphics. I'm talking about computers that do "real work" that is seriously floating point intensive. Far beyond what you'd find running on a small server.

                Get that small blade server out of your head. Think big iron. Think supercomputers. Think racks of cards

                • integrated GPU's don't really come anywhere near the flops of a discreet card (I don't get why people call it discreet instead of discrete, I always assumed that was the proper spelling for the term, anyway). Yes, it's better than the processor by itself, but no it is not substantial. GPU clusters already exist, remember the one from china? One single GPU can perform about 20x as much as a single xeon's GPU pretty easily. I understand what's needed for science research and understand supercomputer clusters

                  • You really should read up on the new generation of integrated GPUs. They have come a long way in just the last year. AMD has integrated a full up ATI 6xxx GPU on die. Intel is making remarkable strides as well.

                    Having a discrete graphics card hanging on a PCIe bus may have been 20x faster last year. Integrated graphics were really inadequate when they were in the northbridge. Again, that was last year. Things have evolved significantly since then.

                    • Is this a joke? I do read up on this continually, and have a 6970 at home and an E350 netbook. Not commenting on my work. It's not a "6xxx" GPU, it's equivalent to the lowest end of the 6 series gpu []. The highest end version is 1/10th that of the flagship graphics card and has only better performance than the lowest end of the flagship design from the last iteration. Also, that hasn't changed *since* at least the last two iterations (we're talking 4x series). I don't know what planet you're on but FLOPS do

              • Oh, and I forgot one more thing. The newest integrated CPU/GPU from AMD uses the same processor core that goes into the HD6xxx series chips coming from ATI. So, contrary to what you say, the FLOPS are not ". . . shit in comparison to any discreet [sic] graphics card . . ."

              • Single discreet GPU does more work than over 100 dual core & quad core CPUs for the workload i need them for ...

                Integrated GPUs still did about 15 CPUs worth of work for that workload ... That is significant, no matter how you put it.

          • by Desler ( 1608317 )

            You do realize that a lot of supercomputing clusters use GPUs for their number crunching, right? So how exactly is it a non sequitur?

            • GPU clusters and GPU's embedded into processors are not the same. Desktops might have iGPU's but server processors don't. Servers that do GPU clusters use discreet graphics cards, commonly. Why do people associate the two? Do people not really understand that on an 8 processor chip if you have a single integrated GPU (let alone 1 per core) its performance is still going to be shit in comparison to even a cheap discreet graphics card?

    • While our HPC cluster core count has increased by a factor of 4 (allowing researchers to do more work), the amount of energy and floor space required did not increase that much at all.

      How much did it increase, then? Just curious about the efficiency...

      • by Enry ( 630 )

        Hard to say. We were already moving to blade servers when we started the expansion. With the previous chassis servers with 90W CPUs, we'd have to get a rack rated at 30KW rather than the standard 20. With the low power CPUs, we can easily fit in a 20KW rack. Our data center folk (who really know the numbers) started to panic when we had a 20KW rack 1/2 full of 1U systems.

    • In addition to this, Google runs DC power supplies, with a low-voltage on board battery as opposed to large rack UPS. I've heard they have some innovative tricks for server room cooling as well, but I've never seen confirmation of exactly what they're doing. But Google goes to great lengths to cut down data center power usage.

    • There are reports that Google has been testing servers [] using low-power many-core servers from Tilera and Quanta. Facebook is also test-driving Tilera chips [] and seeing promising results when using them on key-value pair apps like memcached. When you have 900,000 servers, you get plenty of attention from processor and server vendors.
      • by Enry ( 630 )

        Not all 900k servers are being used for memcached. You will need higher speed CPUs for crawling, operating all the back-end Google services, transcription, etc.

        • by JamesP ( 688957 )

          I'm guessing for IO bound servers (that is, all those that take care of storage) the use of a fast CPU is a waste (unless they're also running MapReduce)

          Since most modern CPUs can 'go around the world' while the HD is fetching data, kind of makes sense.

          Of course, the cost/benefit analysis is not only this.

  • 900,000 servers, and they are all data-mining the internet for porn. Awesome.
  • by StripedCow ( 776465 ) on Monday August 01, 2011 @11:12AM (#36947788)

    With the current pace of technology, those machines will be outdated in a few years.

    Imagine the pile of garbage that will create...

    • by jedidiah ( 1196 )

      Send it to the local PC recycler.

      We have one of those in rural Texas. Surely there are some in the same neck of the woods where Google operates.

      • by Pieroxy ( 222434 )

        Send it to the local PC recycler.

        Should we send the 900,000 units in one shipment?

      • by Desler ( 1608317 )

        Send it to the local PC recycler.

        That may not be the good idea you think it is. []

      • by cpghost ( 719344 )

        Send it to the local PC recycler.

        It's actually a shame that perfectly working machines are being destroyed this way... while in Asia, Africa etc. people (schools e.g.) would be more than happy to use those machines for 5 to 10 more years, at least.

        • Schools in the USA and Europe would be grateful for them too! Businesses often throw out machines as part of a 3-year rolling upgrade cycle, while schools are stuck with machines 5+ years old because they don't have budget for new ones.
        • sure, but what do citizens like you and me do about it? close to nothing I'm afraid. Isn't there about a billion pcs operating world-wide today? what do we do to recycle a few hundred million of them each year?
    • With the current pace of technology, those machines will be outdated in a few months .

      Imagine the pile of garbage that will create...

      • by jdgeorge ( 18767 )

        Maybe more to the point, they'll be unable to fulfill their mission for Google within a few years. They'll be good enough for a while, though.

        • by abarrow ( 117740 )

          So, why is that, exactly? Google has proven by it's actions that their solution to the need for more processor power is just to add more servers. Granted, it'd be tough to pull one of the old blades and play the most recent edition of Duke Nukem, but really, these systems will still be able to crunch numbers for a very long time. Would buying faster systems with more cores would allow a single system to crunch more? Sure, but really, those old systems can still happily serve their original purpose. As Googl

          • Not exactly. When cost of not replacing hardware (in increased power consumption, rack space, cooling requirement etc) exceeds the cost of replacing hardware ; they replace. Even when the trashed hardware doesn't "actually break".

    • I expect quite a high percentage of that to be recycled, actually.

      • Not sure were you work, but in all places I've worked hardware was stockpiled once it was put out of service. I'm not so sure about the reason, but there seems to be a lot of accounting issues for a company if it wants to get rid of stuff. Maybe someone more knowledgeable can comment on that.

    • No, people like us will buy them for our home networks. Low-end and low-income users will happily use them. Old machines have *a lot* of life, maybe not on the front lines anymore.

      • I was able to recycle a dell inspiron 8500 back in 2005. Just died on Friday, 6 years working well, the drive finally crashed. I figured I would fix it and give it to a non profit that might be able to use it.

        • by swb ( 14022 )

          A lot of non-profits won't take donated systems anymore because it's a nuisance to deal with so many antiquated systems. Non-profits need working and reasonably contemporary systems to do their work, a bunch of 256 meg Win 98 systems is really more of an insult than a benefit.

          • Really? then i guess I'll fix it up for myself and use it as a travel computer. Thank you for advising me.

      • by Desler ( 1608317 )

        Low-end and low-income users will happily use them.

        No, they would more likely rather having some recent that actually works well rather than some 12 year old dumpster dived computer that runs as slow as molasses.

    • by LWATCDR ( 28044 )

      Well that will depend on a lot of things. If a socket upgrade is available they can just put in a new CPU. Even if the "server" is taken out of service you are just talking about the mainboard, CPU, and memory. The CPU and memory might be offered for sale as "used" or "refurbished". The board will be recycled, the power supply and rest will probably be reused unless a more efficient solution is available. Servers tend to have a longer life than say cell phones, desktops, and PCs.

    • by Sitrix ( 973862 )
      I am quite sure that not all of it is what we call "physical" servers. It's most likely a cluster of beefy hardware running a ton of VM's. As that hardware becomes obsolete, engineers will run less VM's on it and later move it out of main production environment to handle less stressful tasks. It's common now, seeing several servers (48Cores, 512GB ram) running a few hundred virtual servers. So it will take a long time before that hardware will be completely thrown away...
      • by Quila ( 201335 ) on Monday August 01, 2011 @01:46PM (#36949838)

        Dual-processor, two SATA hard drives, 12V PSU, 12V Lithium battery. It's not even sealed in a case, just a frame holding a board, with the PSU, battery and hard drives held on with Velcro.

        Most of these will be about that spec.

      • I am quite sure that not all of it is what we call "physical" servers.

        Google has a different model. Quila mentioned what their hardware is like; here is a slightly outdated Wikipedia article describing it []. Google would prefer to avoid the overhead associated with running virtual machines. 10-20% overhead may not be a lot for an organization with smaller computing needs, but with Google that would mean adding another 90,000-180,000 servers. Their computing needs are way beyond what any individual server can do anyway.

        What they've been doing instead is writing their apps for

    • Did you ever see the video about how Google disposes of/recycles their hard drives? (Video Google 'google data center')

      I'm quite sure they have a protocal for how they recycle all of their hardware.

    • You say this when there are people using Commodore 64s as web servers.

      I think the machines will find a spot somewhere. That, or they're mostly recycled.

  • Impressed (Score:4, Funny)

    by Verdatum ( 1257828 ) on Monday August 01, 2011 @11:13AM (#36947806)
    No comment about it being over 9000 yet. I'm impressed Slashdot.
    • by saihung ( 19097 )

      You absolute bastard.

    • I'm not as impressed with that as I am with the fact that no one ha yet pointed out

      data centers are using less energy that projected

      . "Data centers that projected what?", you might ask, as I did. Nope, its a typo. Or an omission, since data centers certainly could project... but more likely a type.

      • by Hooya ( 518216 )

        "but more likely a type" what?, you might ask, as I did. Nope, it's a typo. Or an omission, since you could have a type of grammatical error... but more likely a typo.

    • by Nidi62 ( 1525137 )

      Don't worry, eventually there will be at least 9000 comments all with some variation of "But that's over 9000!"

      Including this one

  • Couldn't resist.
  • by instagib ( 879544 ) on Monday August 01, 2011 @11:15AM (#36947854)

    I wonder how a few hundred mainframes plus storage arrays would fare in terms of TCO.

  • OT Google is using all kinds of renewable sources for their energy. []
    Back OT Do you think they keep all their servers in mobile homes so they can keep the number of servers a secret?
    • by Nidi62 ( 1525137 )
      Seems like they always keep what's in their locations a secret. My father was a manager at a distribution center for a fairly large national electrical supply chain, and several times people would come in to buy things for a complex they were building nearby. Apparently they worked for Google (they were always wearing Google shirts) and they were never allowed to tell them what they were building or what kind of work they were going to be doing.
  • by roman_mir ( 125474 ) on Monday August 01, 2011 @11:20AM (#36947924) Homepage Journal

    So does it have enough data to answer the last question meaningfully yet?

  • Noticed that the gentleman in the picture is wearing fullsize earphones or ear protection. Is the room that noisy, or is he just enjoying some tunes?
    • by Bengie ( 1121981 )

      Google had a YouTube video of their security practices. They do actually have hearing protection for the server rooms.

  • Virtualization is very inefficient compared to simply running multiple server processes on a single box, because each VM allocates resources to an instance of the OS, and RAM is more-or-less statically allocated beetween them. This makes sense when running several different services that each require a different operating environment, or to enforce complete user separation, e.g. a hosting service. But I would imagine google is running tens of thousands of identical servers running the same server daemon,
    • by CBravo ( 35450 )

      Young padawan, there are many sorts of virtualization.

    • by dmpot ( 1708950 )

      I would imagine google is running tens of thousands of identical servers running the same server daemon, so why would Virtualization make sense and save energy there?

      Who said that Google uses virtualization to run identical servers?

      Just running "git log --grep=virtualization" on the Linux kernel, you can see that Google does not contributed much to virtualization in the Linux lernel, in sharp contrast to other part of the kernel such as ext4.

  • beowulf cluster of these?

  • ... that's a square 1000 x 1000 meters. Now place 5 "normal" computers next to each other in two layers and you need about 100 000 square meters. Divide 100 000 by 8 data centers (Atlanta, North/South Carolinas, Chicago, California, Oregon, Taiwan, Ireland) = 12 500 square meters per data center. If every data center has 2 floors than you need a building like 80 x 80 x 5 meters. And you still have enough place for the guys with wheelbarrows :) Anyway a data center like this would be about a size of an indus

"I prefer the blunted cudgels of the followers of the Serpent God." -- Sean Doran the Younger