Slashdot Mirror

← Back to Stories (view on slashdot.org)

Report: Amazon Cloud Backed By 450,000 Servers

Posted by ryuzaki0 on from the single-breasted-cumulus-spotted-in-lorton dept.

1sockchuck writes "How many servers does it take to power Amazon's huge cloud computing operation? A researcher estimates that Amazon Web Services is using at least 454,400 servers in seven data center hubs around the globe. The analysis suggests up to 70 percent of those servers may be in Virginia."

14 of 45 comments (clear)

  1. Obligatory xkcd by zill · · Score: 4, Funny

    Wrong, it's actually just one machine.

  2. Bad Math and Bullshit Story by Anonymous Coward · · Score: 5, Insightful

    So, first off, the guy who came up with this number made a ridiculous number of assumptions with no real evidence to back them up, so the number is completely meaningless. Also, from TFA:

    Liu then applied an assumption of 64 blade servers per rack – four 10U chassis, each holding eight blades – to arrive at the estimate.

    Now, I might have to go dig out my TI-82 to doublecheck, but I think I see a small flaw in this math.

  3. Re:Go Virginia by cashman73 · · Score: 5, Funny

    Except I suppose the citizens of Virginia aren't going to be too happy when the government realizes that Amazon has a "physical presence" in the state and decides to start charging them sales tax,. . . D'oh!

  4. Just as important... by jtseng · · Score: 4, Insightful

    What about their support infrastructure? I don't care about the physical locations, but I'm wondering about how many UPS banks do they have? How many primary power feeds do they have to each location? How long do the diesel generators last? Electrical transformers? As a customer, I'm not just concerned about scalability and capability - I want to make sure my presence is always available too!

    --

    Sanity.html - Error 404 not found

    1. Re:Just as important... by Anonymous Coward · · Score: 2, Interesting

      Having been to some of the largest data centers in Virginia, I can assure you that the UPS warehousing is quite vast as well - certainly not an afterthought like in many data centers.

    2. Re:Just as important... by bobcat7677 · · Score: 2

      Except a major advantage you are *supposed* to have in the cloud is geographical diversity. So if a major disaster strikes in one area and turns one massive data center into a smoking crater (worst case) or simply some boneheads deploy faulty code to all the cloud servers in one location (best case), then there are servers elsewhere that can take up the slack. If there is too much concentration of servers in one geographical location, and that location goes down, then effectively the cloud is down even though that should not really ever happen short of a global disaster.

    3. Re:Just as important... by ILongForDarkness · · Score: 3, Interesting

      Exactly. If 70% of the capacity is in Virginia can the other 30% keep everyone going? The other locations probably have a lot of load too because people tend to have instances in different zones to try to route traffic to nearby clusters for example. So lets say europe and asia each have the other 15% (I realize there is some other datacentres in the US so not exactly true but approximate). You have 2B+ people in asia all hitting that 15% of hardware already. Now you try to shovel off 35% of your network to their cluster ie greater than 2X more work and have all the extra latency issues to deal with (because presumably US instances were being used because they were closer to the users) ouch.

      That said though I guess we don't know how their usage looks. They might be double sized already and so it would only be 0.5*(15+35)%. I doubt they are keeping that much spare capacity around but who knows?

  5. what are these server things? by alen · · Score: 4, Funny

    i thought the cloud was this magical circle in a white paper where all the data just lives

  6. Re:Go Virginia by Anonymous Coward · · Score: 3, Informative

    Too late. Starts next year, apparently. :(

    http://hamptonroads.com/2012/02/virginia-sales-tax-looms-amazoncom

  7. Article is bogus. by Barbara,+not+Barbie · · Score: 2

    I down-rated this in the firehose - it's all guesswork. I know "pull numbers out of your rectum and get page hits from slashdot, because slashdot is the new Mikey - they'll post anything!"

    --
    Let's call it what it is, Anti-Social Media.
  8. Re:Err. What? by PTBarnum · · Score: 2

    So Amazon built a 17,000 core supercomputer. That just means they were able to allocate 17,000 cores to a supercomputer project and run benchmarks on them. Presumably the vast majority of the fleet was still serving external customers. I'm fairly confident that Amazon is not running benchmarks on cores which are currently being used by customers.

  9. Re:Go Virginia by PTBarnum · · Score: 2

    Virginia, or more precisely the DC suburbs near Dulles Airport, is a popular spot for data centers for the same reasons as the bay area. Lots of companies have data centers there, so there is lots of infrastructure, so lots of companies have data centers there, so .there is lots of infrastructure...

    I think the TLAs you are thinking of are more than capable of interfering with data centers anywhere in the country.

  10. My Version by twmcneil · · Score: 2

    I'll guess an even 1,000,000 servers give or take. I could be off but 450,000 is way low I think. But... my estimate has the same accuracy as TFA.

    --
    "The ferrets, they're every where I tell you!"
  11. um by ILongForDarkness · · Score: 4, Insightful

    No still a guess. From the blog he says he can only discover a rack if he managed to get a instance on it. So yes there might be racks that he never sees (for example I'm sure Amazon reserves some racks for themselves) but also he is assuming that the rack is full if he sees it. As well he is assuming he is right that the networking is done on a per rack manner for all the datacentres. Who knows different datacentres might do it differently (for example maybe europe only has half the servers of a US based datacentre but to keep the number of vlans the same they split the racks in half and only use the first half of the /22 IPs, maybe Amazon has a crap load of racks half full because they haven't gotten around to installing all the equipment, are in the middle of a hardware refresh, debating on having NAT or compute chassis in the available space etc. The only way to have a reasonable idea is to knock on the door and ask them. If they answer they "might" be telling you the truth But "researching" from the outside? You have know idea what you are looking at.