Slashdot Mirror

← Back to Stories (view on slashdot.org)

Intel Confirms Decline of Server Giants

Posted by Soulskill on from the bigger-they-are-the-cloudier-they-fall dept.

An anonymous reader writes "A Wired article discusses the relative decline of Dell, HP, and IBM in the server market over the past few years. Whereas those three companies once provided 75% of Intel's server chip revenue, those revenues are now split between the big three and five other companies as well. Google is fifth on the list. 'It's the big web players that are moving away from the HPs and the Dells, and most of these same companies offer large "cloud" services that let other businesses run their operations without purchasing servers in the first place. To be sure, as the market shifts, HP, Dell, and IBM are working to reinvent themselves. Dell, for instance, launched a new business unit dedicated to building custom gear for the big web players — Dell Data Center Services — and all these outfits are now offering their own cloud services. But the tide is against them.'"

17 of 152 comments (clear)

  1. If Google sold servers... by denis-The-menace · · Score: 5, Interesting

    If Google sold servers, HP and Dell would die overnight.

    Just the "12volt-only" power supplies with built-in batteries with "12volt-only" motherboards makes them more reliable than anything out there.

    HP and Dell either can't or won't license this from Google.

    --
    Obama's legacy: (N)othing (S)ecure (A)nywhere and (T)error (S)imulation (A)dministration
    1. Re:If Google sold servers... by denis-The-menace · · Score: 3, Informative

      In google servers, the power supply only make +12volts.

      There are no -12V, +5V or -5V rails.
      There are, instead, DC-to-DC converters on the motherboard.

      --
      Obama's legacy: (N)othing (S)ecure (A)nywhere and (T)error (S)imulation (A)dministration
    2. Re:If Google sold servers... by postbigbang · · Score: 3, Insightful

      Other motherboards make use of similar dc-dc converters and have for a long time. It's nice to have a 12vdc bus; makes it more dense. But it's neither innovative or unique. Instead, it's all about density and design for a specific purpose. These aren't retail-able machines. And there are now luscious racks you can obtain with lots of dense Intel, AMD, and even ARM-powered systems. If you have the application, someone has a design.

      It might be a good design for you, and not for others.

      --
      ---- Teach Peace. It's Cheaper Than War.
    3. Re:If Google sold servers... by fm6 · · Score: 4, Interesting

      Sorry, you're wrong. Wish you were right.

      I've always been appalled by the way PCs rely on big, hot, wasteful noisy internal power supplies. When IBM entered the workstation market, 30 years ago (Oh, Lord, that makes me feel old) I worked for a company that made a pre-PC x86 system that relied entirely on external, passively cooled power supplies. To me, this was clearly the way of the future, but once IBM entered the market, everything had to be IBM compatible, even the way the power system worked. Because if you couldn't use IBM-compatible power supplies, your system cost too much to build. (I once had to throw out a perfectly good Zenith PC with a blown PS; although it was mostly IBM-compatible, its power supply was proprietary, and cost too much to replace.)

      So, Google can't go into the hardware business, because their machines would cost too much and would rely too much on proprietary infrastructure. Easier to justify using your own technology regardless of cost when you're gigantic and profitable.

      HP and Dell's nightmare isn't Google. It's cloud computing in general. The cloud providers (which includes Google, if you ignore the fact that they only provide high-level cloud services, unlike Amazon) mostly build their own hardware. Those that don't buy cheap no-name hardware.

    4. Re:If Google sold servers... by afidel · · Score: 3, Interesting

      Not in the least bit, Google designs their servers to optimize power usage and absolute lowest cost per compute cycle. Those are not the same goals for every server buyer. For instance single threaded performance is a large factor for me because we run a lot of interactive workloads that are single threaded or weakly threaded but Google doesn't really care about single threaded performance because they're optimizing at the datacenter level. I also care a lot more about the reliability of any given unit because my jobs are mostly traditional single-server jobs with only my most critical workloads being clustered so the loss of any given node has a significant impact on my overall reliability whereas Google can lose dozens of servers a day per datacenter and it would have no impact on their overall operations. Another example is storage, Google uses COTS SATA drives with horrible MTBF stats and they do so without RAID protection, the only application where that might remotely have a chance of working for me is Exchange 2010 because I have four copies of each database online and the client is seamlessly pointed to a working copy.

      --
      There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
    5. Re:If Google sold servers... by labradore · · Score: 3, Interesting

      What's the point?
      1. you use less parts and cheaper parts in the power supply.
      2. you have fewer and shorter cables
      3. you use 5V, 3.3V, regulators that are the right size for the job. this saves space and saves material
      4. you get to choose where to put these regulators so that heat management can be more optimal
      5. it's easier to integrate the 12v battery with the space saved

    6. Re:If Google sold servers... by drsmithy · · Score: 3, Insightful

      Your statement about BGP makes no sense to me. How does BGP interfere with cloud-type connections and not others?

      He is rehashing - in a rather rather pained and circuitous fashion - the "if you lose your internet connectivity you can't do any work" argument.

      This point is not entirely without merit, but generally fails to recognise that a) most companies these days can't do a lot of work without an internet connection anyway and b) internet connectivity is usually a lot easier and cheaper to make highly available and redundant that server infrastructure.

  2. Your first server, in 2012 by Compaqt · · Score: 3, Informative

    Back in the day (say, 2008 as in the article), if you wanted to buy a server, you'd buy one from the big three.

    These days, especially with FB and Google leading the way on commodity hardware, it's a different story.

    So what should you get for your first server. I.e., you're a small company. You've got a couple of laptops. You're outgrowing mutual Samba.

    You maybe want a fileserver. Maybe it'll have a few NICs and a virtual machine on it (Xen?) will do double duty as a external webserver.

    So, Core i3, i5, Xeon? Number of processor cores? Forget fast drives, and just buy a lot of memory? Rack? Or tower?

    Lockable front (so people can't just come by and reset it)? Hotplug harddrives? (You don't go this if you go the Google build-your-own route.) Redundant hard drives and ECC memory? Or a couple different commodity-style servers + sharding/rsync?

    Is a big 3 server worth it? Or search for your own server case + server power supply, etc.?

    --
    I'm not a lawyer, but I play one on the Internet. Blog
    1. Re:Your first server, in 2012 by Anonymous Coward · · Score: 3, Informative

      Search for your own. Priced one from hp/dell and it would have cost $6,000 plus. Built it with the same specs for $3000. That right there is why their server sales are dwindling.

    2. Re:Your first server, in 2012 by ard · · Score: 3, Insightful

      With the same specs? With hot-plug drives, true hardware raid, iLO/iDRAC lights-out management, secondary bios if flashing fails?

      Get a refurbished HP gen 5 or 6 server instead of building your own. Perfomance will be sufficient, don't worry. It's well below $3000, and you get enterprise quality hardware.

    3. Re:Your first server, in 2012 by MightyMartian · · Score: 4, Insightful

      As much as anything, I think virtualization is murdering the market. I bought a $3000 server that hosts six VM guests; two Windows installs (one a DC, one an Exchange server) and four Linux. A couple of years ago, I would have needed at least three servers to do it (one for each Windows install) and one Linux. Admittedly they wouldn't have to have the balls that the new server has, but still, I think we'd be talking about $4000 to $6000 in hardware. Even worse, these are all just basically images sitting on hard drives, so they can essentially be perpetual. Two or three years, when the current server dies or I decide I need more juice, just move the VM images over and away I go, and with hardware prices the way they are, I doubt the next generation server will cost any more than the one I have now, and maybe even less.

      Factor in the cloud, VPS hosting and so on, the demand for servers will inevitably drop.

      --
      The world's burning. Moped Jesus spotted on I50. Details at 11.
    4. Re:Your first server, in 2012 by drsmithy · · Score: 3, Insightful

      Is a big 3 server worth it?

      Almost certainly. The problem is most techies - especially young ones - only look at a handful of specifications (CPU, RAM, # disks) and the sticker price, because they think their time is free.

    5. Re:Your first server, in 2012 by h4rr4r · · Score: 4, Interesting

      Linux software raid is great.
      Proprietary software raid is garbage.

      I base this on what I have seen. Linux software raid beats cheapy hardware controllers both in reliability and speed.

  3. No surprise. by heypete · · Score: 5, Interesting

    Why bother with branded parts made by an ODM when you can buy directly from the ODM?

    My old workplace had (has, probably) a fairly beefy Sun server with a whole bunch of disks. They used it as a RAID-based storage server for a bunch of lab data. As they do on occasion, a hard disk would crap out. The server wouldn't take ordinary disks, though: it would only accept Western Digital disks with some Sun ID code baked into the firmware -- rather than simply being able to buy a few WD RAID-friendly disks ahead of time, we had to jump through Sun's hoops to get disks replaced under warranty. This usually was a multi-week process, during the array with the failed disk was running with a hot spare -- hardly ideal. That was the last time we bought Sun systems.

    At some other point, we were planning on setting up a few more storage servers for backup data. Dell's price for a storage system, including firmware-locked drives, was about triple the cost of doing it ourselves with SuperMicro servers, MD-based software RAID, and RAID-friendly disks. We ended up buying two of the SuperMicro-based systems and putting them in different buildings for semi-offsite backup (the concern was if the server room caught fire, not if a meteor affected the whole city). The only extra step during the setup was putting the disks in their caddies: the Dell systems came with the disks pre-installed. That took about 5 minutes per server. Whoop-dee-doo.

    The Dell servers restricted our (with firmware-locked disks) options and cost substantially more than doing it in-house. We'd be stupid to go with their products, as we'd be locked to that vendor for the life of the servers.

    Sure, we had Dell Optiplex systems as the desktop workstations for researchers as they were inexpensive, reliable in the lab, and essentially identical (useful for restoring system images from one computer to another), but their server stuff is stupidly overpriced.

    The SuperMicro servers were much more "open" in that they used pretty bog-standard parts and didn't have stupid anti-features like firmware locking.

    1. Re:No surprise. by heypete · · Score: 3, Informative

      First, a RAID array does not "[run] with a hotspare." When a failure occurs, the hotspare becomes a fully integrated member of the array, at which point you would be running without a hotspare, which on a redundant array isn't that much of a problem considering the Dell replacement would be there within 4 hours of reporting/determining a hardware failure.

      It took Sun 3+ weeks to send us a replacement hard disk under warranty and required multiple phone calls. This happened on multiple occasions and was one of the main reasons we decided to stop buying Sun servers.

      Yes, the spare became an integrated member of the array. That's true. My point was that the hot spare was now a member of the array and we had no remaining spare disks in the array. Since the server hardware only allowed drives with the Sun firmware, we couldn't keep a supply of spare disks around to swap into the arrays as needed.

      Second, Dell servers do not have "firmware-locked disks." I've never heard of such a thing. It's a pretty absurd concept that you could only have OEM hard disks in your box, and an unrealistic expectation that clients would comply.

      They did: "In the case of Dell's PERC RAID controllers, we began informing customers when a non-Dell drive was detected with the introduction of PERC5 RAID controllers in early 2006. With the introduction of the PERC H700/H800 controllers, we began enabling only the use of Dell qualified drives."

      Same thing with Sun, at least at that point in time.

      Finally, hardware RAID is leaps and bounds above software RAID. There's a reason it's cheaper to go with software...

      Software RAID was perfectly adequate for our needs: as backup servers they didn't need to have the utmost performance. As a bonus, we weren't reliant on a specific make and model of hardware RAID card: we could connect the array to any system running MD. Even under heavy load the demand on the CPU was negligible.

      The Sun server was the main Samba share for the lab: lab instruments would write data to it and researchers would access that data on their desktops. It also used software RAID with multiple arrays set up. CPU usage was similarly low, even at high loads, and it worked quite satisfactorily for the lab.

      You might have saved money up front, but over the life of the server, you could potentially lose much more when you consider catastrophic hardware failure which would be fully covered under the warranty of the Dell box.

      SuperMicro offered a comparable warranty, so that wasn't really an issue.

  4. FIFTH? by Anonymous Coward · · Score: 3, Interesting

    Let me get this right. Google, who builds all of their servers in-house, exclusively for their own use (not for resale), is the fifth largest buyer of Intel server chips in the world?

    That sure paints a picture about the sheer size of Google's data center operations.

  5. Low- vs. High-level cloud services by DragonWriter · · Score: 3, Informative

    The cloud providers (which includes Google, if you ignore the fact that they only provide high-level cloud services, unlike Amazon) mostly build their own hardware.

    Google provides low-level cloud services (IaaS in the form of Google Compute Engine, PaaS in the form of Google App Engine, RDBMS-in-the-cloud in the form of Google Cloud SQL, bucket-style storage in Google Cloud Storage) as well as higher-level services (all of Google's various apps build on their cloud infrastructure.)

    So the Google-Amazon distinction drawn in the parenthetical is inaccurate.