Surveying the World of the Biggest Server Farms 106
1sockchuck writes "Rackspace said this week that it is managing more than 50,000 web servers, raising the question: who else has that many? Of companies that publicly discuss their server counts, there are only a handful that are near or above the 50,000 server mark, including 1&1 Internet, The Planet, and Akamai, as well as Rackspace. The larger totals are found among companies that don't discuss how many servers they're running. The leading suspects: Google, Microsoft, Amazon and eBay."
Shouldn't it be easy to figure out? (Score:3, Interesting)
Re: (Score:3, Informative)
Re: (Score:3, Interesting)
Re:Shouldn't it be easy to figure out? (Score:5, Interesting)
A lot of datacenters get built in repurposed buildings - the square-footage is often misleading (some are even 60's era compute farm housing - 90% of the space may be unused.)
For low-latency datacenters, you build in the middle of cities. Then you find square-footage really doesn't cover it: the fire-marshall shows up and red-tags you because he doesn't want a six mega-watt dense power sink in the middle of his premium real-estate.
Re:Shouldn't it be easy to figure out? (Score:4, Insightful)
Also hopefully they are not counting virtual machines here.
--
Slow Poke [pair.com]
Re: (Score:2, Interesting)
Also hopefully they are not counting virtual machines here.
--
Slow Poke [pair.com]
It's almost guaranteed that they are talking virtual servers as well as some bare metal. Rackspace is talking about the 50,000 web servers that *THEY* manage--i.e. managed services. This figure likely doesn't include the thousands of rented colo machines which their customers manage themselves. The latter was Rackspace's first business model, a pure colo. It wasn't until relatively recently that they started offering managed services. It may very very turn out that the 50,000 server figure is the entir
Floor space requirements can change over time. (Score:2)
Keep in mind that companies which use mainframes have seen a fairly large reduction in machine size over the past 20 years or so. Old water-cooled machines and rows of tape drives are being replaced with CMOS machines and tape silos, each taking of a fraction of the floor space that their predecessors did.
When I started working for Northwest Airlines, for example, there were multiple IBM and Unisys mainframes sitting in the main computer center, all residing in a large temperature-controlled raised-floor r
Re:Shouldn't it be easy to figure out? (Score:4, Insightful)
That's going to come up with some off numbers.
There are oddities in there. Like, Google has some decent square footage with QTS in Atlanta.
Is the number, the number of machines in house, or the number of machines managed by the company in house? Places like Equinix have huge facilities, but they don't manage them (except for helping hands support).
Just because a company takes a bigger footprint, does that make them bigger? My old place had an Alexa rank in the top 300, but the main sites were served from maybe 24 machines set up for well over 10 million users/day, not including hot spares and ancillary equipment (DNS, mail, etc). That was unique users, not requests or page views. :)
I know Quantcast has a huge footprint, but they're in someone elses facility. I think it was only one of a few DC's that they're in. I didn't know it was their equipment until I talked to one of their techs. It was a very nice setup. The conversation with my coworkers went "I want to set our stuff up like theirs. Too bad we have dissimilar equipment, it'll never look so good."
Depending on who you're looking at, the footprint isn't the front end either. Places like Google, Quantcast, and many others have a LOT of non-public equipment for doing the fun stuff. Like Google has a slew of spiders crawling, analyzing, etc. It's perfectly likely the could (not necessarily do) get away with just a few dozen front end machines for Google.com, and pass the work off to back end machines.
It's all in the strategy that they use. If you're a multibillion dollar operation, do you squeeze every bit of power you have out of a machine, or stay real low and spread it across many? For my old place, I set up to squeeze everything I could out of them, and then spread out across machines so we could lose machines (hardware failures usually) without hurting the site. We didn't have the budget (the bosses liked the profit), but if I had really wanted, I could have probably spread it across thousands of machines. It just makes for headaches and larger IT staffs to keep up with it.
Re: (Score:2, Interesting)
Re: (Score:2)
Re: (Score:1)
Depends...do you count raw tonnage of servers, or do you include the ancillaries like cable runs, UPS, cooling etc?
Easy way to win:
My Eniac replica, combined with my replicas of Mayan and Egyptian pyramids (purportedly used as astronomical computers...you know, by the illuminati, etc ;)) means that I win by sheer tonnage!
50,000 web servers, not physical servers (Score:3, Insightful)
They're using Netcraft to prove their server count - which reports on IP addresses. Just because there are 50,000 IP addresses responding to port 80, doesn't mean they have 50,000 boxes. The shared hosting arrangements can easily have dozens and dozens of "servers" operating on the same physical box.
Yes, it's still impressive... but not as impressive as it would first appear.
Re:50,000 web servers, not physical servers (Score:4, Insightful)
They're using Netcraft to prove their server count - which reports on IP addresses. Just because there are 50,000 IP addresses responding to port 80, doesn't mean they have 50,000 boxes. The shared hosting arrangements can easily have dozens and dozens of "servers" operating on the same physical box.
Not exactly what netcraft is saying right now [netcraft.com] from TFAL:
Netcraft has developed a technique for identifying the number of computers (rather than IP addresses) acting as web servers on the internet, and attributes these computers to hosting locations through reverse DNS lookups.
I suspect there's a power law in here, and that the estimates of google and amazon and others should be way beyond this, perhaps surpassing 500.000. That would be an interesting milestone.
FTFA:
There's a widely circulated estimate of 450,000 servers, but that number is at least three years old. If it was ever accurate, it certainly isn't anymore, given Google's data center building spree. Google's recently revealed container data center holds more than 45,000 servers, and that's a single facility built in 2005.
Re: (Score:3, Interesting)
4 years ago. I wonder if all those boxes are still running right now? I wonder what google does when it retires servers.....it would be kind of cool to have a couple of bonafide google racks doing something cool at my house.
Re:50,000 web servers, not physical servers (Score:5, Informative)
If this [google.com] is to be believed, they salvage and rebuild as much as they can, and send the rest to recycling.
Re: (Score:2)
Plenty of years ago I was in a large datacenter, where Google had a cage full of their racks. These were the old DIY "4 MBs on a tray" systems.
>
>it would be kind of cool to have a couple of bonafide google racks doing something cool at my house
>
Let me tell you something, you don't want to have those servers in your basement, that setup definitely wasn't cool :) the amount of heat that came out of that row of racks was pretty enormous.
OK, this was long time ago, I expect that Google's setup has chan
Re:50,000 web servers, not physical servers (Score:5, Informative)
If you bother to RTFA you'll see it isn't a netcraft server count, but a mention in their SEC reported earnings something they are unlikely to just make up.
The comparisons are with netcraft numbers. And those netcraft numbers are explicitly not IP address counts and have rackspace as way under 50,000 (which you would expect since many machines wouldn't be web servers (database backends, mail servers, etc) and many would be firewalled to not allow public access.
Re: (Score:2)
The shared hosting arrangements can easily have dozens and dozens of "servers" operating on the same physical box.
Dozens?! I worked at a web host like 8 years ago and we were running up to 400 servers shared on a single box. I assume you can squeeze on a hell of a lot more these days!
Re: (Score:1)
Re: (Score:2)
You bring up the magic question, how many virtual IP's can be on one box? I don't know the limit on a Linux machine. I brought up an entire private /24 on one, just for the fun of it. So, if you considered that each of 50,000 machines had 254 IP's on them (a wild idea, but still), they could have just 197 machines.
Re: (Score:2)
No, they're using Netcraft for the companies that didn't release the information.
TFA:
Here's a look at some of the providers with high server counts, gleaned from public reports and partial data from a recent Netcraft server count report
There's a little note after each company where the data came from, all of the top 5 came from the company itself.
Re: (Score:1)
analogy (Score:1)
Re:analogy (Score:5, Funny)
Are you trying to show who has the biggest dick in the IT world ? :3
No, didn't you read the summary? We're trying to find out who has the biggest rack.
Re: (Score:2)
Cowboy Neal???
Re: (Score:1)
We're trying to find out who has the biggest rack.
Bullwinkle!
Who else? ME. (Score:1)
Thats right, me. In fact I just set up #50,001. You'll never guess what I'm doing with them either.
Re: (Score:3, Funny)
Thats right, me. In fact I just set up #50,001. You'll never guess what I'm doing with them either.
Do you really need that much porn?
Re:Who else? ME. (Score:5, Funny)
Re:Who else? ME. (Score:5, Funny)
Re: (Score:2, Funny)
Re: (Score:2)
So does that make you feel like you're REALLY an evil overlord in a secret lair?
(reference to a dilbert cartoon about getting a second monitor)
Keyword: Evil (Score:2, Insightful)
Re: (Score:2)
Probably big server farms, like the ones of google, yahoo, amazon, etc are orders of magnitude over that. There are estimates of 200k servers in 2005, or 450k in 2006 (according to Wikipedia [wikipedia.org]). By now could be the order of millons.
Re: (Score:3, Insightful)
> By now could be the order of millons.
Forget about "could be". According to IDC (feb 2009):
"According to IDC's Worldwide Quarterly Server Tracker, [...] 2008 worldwide unit shipments grew 2.0% to 8.1 million units. [...]" [http://www.idc.com/getdoc.jsp?containerId=prUS21703309]
With 8 million units cranked out in just a single year, there must be 30+ million servers "up and running" in current datacenters.
Which is in sync with my experience with our customers. Ten years ago (which was the time of the dot
Yahoo? (Score:3, Insightful)
Seems to me the second largest search engine likely has 50k servers or more..
Re: (Score:2)
I've always been curious about how Blizzard's WoW servers are spread out around the world.
Re:Yahoo? (Score:4, Informative)
US_realm_list_by_datacenter [wowwiki.com]
Re: (Score:2)
Interesting! Still, I wonder about server configuration, load, and such. It's obviously a good model of dividing work between client and server, but there must still be some times that are quite taxing.
Re: (Score:3, Informative)
Re:just look at the cooling equipment (Score:5, Informative)
Re:just look at the cooling equipment (Score:4, Informative)
Microsoft would have a total pile but since they can't even do SSL on their update sites they are running cheap and probably have less than 300k even with hotmail
The update protocol does winhttps (SSL). The actual file downloads are simple winhttp, since they are signed.
What about porn? (Score:4, Insightful)
From TFA (Score:3, Informative)
Apparently it's not just the number of webservers (just IP addresses), but the number of physical boxes these guys are running. If Netcraft's technique is valid, then it could be helpful in determining the 'true' penetration of FOSS based server installs on the Internet. This could severely impact the ranking of sites that are hosted on certain proprietary OSes.
Re: (Score:2)
They also do not count non-HTTP servers:
The survey does not attempt to count back-end servers (application or database servers) or servers other than web (HTTP) servers.
One more thing: some hosting companies provide private network only servers, not visible outside of the virtual private network assigned to the customer. Perfect for the backend.
Softlayer does that, for example.
If you can count them... (Score:1, Funny)
Obviously, if you can count them, then you don't have all that many.
"A million here, a million there and pretty soon you have a real server farm."
Financial Services Industry (Score:5, Interesting)
Most of them won't go into detail, but Wall Street firms have immense server farms. Some of them are limited in size by the amount of electricity the New York City power grid can supply them. They also have huge data centers in less prime real estate, but microseconds are dollars in the financial markets, so they try to keep as many of their systems as close to the action as possible. There are entire floors of NYC skyscrapers full of racks modeling the financial markets in real time, conducting transactions, and crunching numbers for human analysts.
Re: (Score:2)
Using flawed algorithms and crap assumptions as input. See where that got us.
Re: (Score:3, Insightful)
Using flawed algorithms and crap assumptions as input. See where that got us.
These people don't have "maintaining a stable economy" as a goal. They have "making more money than the other guy". If the market drops more than they do, they consider themselves winners. They can trade into a position that will come up! Who cares what that fluctuation in the market means?
2 CPU servers are a joke (Score:1, Interesting)
Companies that have "real" computers - 32-cpus and more and have 10,000 of those are more impressive to me. Having 100,000 servers, all the same, is impressive, but still will use about the same management as 10,000 servers would.
I've deployed (3) 48-way servers with 2 for HA Oracle DBs and the other for DR and testing. Oracle RAC was the best in class at the time - grid didn't exist. I've deployed hundreds of custom servers (diff OS with diff required patches) running specialized applications from many,
Re: (Score:3, Insightful)
And you seriously think that Google (or any other bigger company) is still using 2-CPU servers?
Re: (Score:1)
If you ask.. (Score:5, Funny)
... Google how many servers it uses, does that mean it's self-aware?
Re: (Score:2)
I am sorry, I can't answer that Dave.
Re: (Score:3, Funny)
Two hosting providers ago... (Score:4, Funny)
I'm pretty sure the hosting company I had a few years ago (aka "kiddie hosting") had that many customers on the server that I was on. Does that count?
Google Data Center Tech (Score:5, Informative)
Hello,
I have worked supporting Google's servers in one of my former employers data center. What I can tell you about there deployments is as follows:
1) 20,000 Servers in our data center; they occupied 8 other sites (~160,000 servers). Our site was one of the largest.
2) Over 30 GigE connections feeding into dual Juniper M20 later upgraded to Juniper T-320
3) Yes they run a custom version of RH
Now for the record; they had approx 160,000 servers in our companies data centers. I have met techs from other data centers which had similar counts. At a minimum I can confirm approx 160,000 and potentially 320,000 and up for other data centers; providing they mirrored their servers. It wouldn't make sense to put all your eggs (servers) in one basket. The time frame for these numbers was back in the early 2002.
Re: (Score:2)
Have you all missed the Slashdot story many months ago, where the reliability of HDDs was discussed with a study from Google?
It said, that at that time, they were adding 100,000 servers a month to their data centers!
They had, I thing 500k of them back then.
Re: (Score:1, Informative)
I am the originator of this thread. Yes, I know for a fact that they have expanded. The contract that was signed with Google was "Right of First Refusal".
The "Right of First Refusal" gave Google the power to be the first to occupy new space before any additional customers. At the time they had 1/3 of the data center. My old co-workers said that they had expanded again over the years.
All I know is that they their deployment is even larger. At this point I have no idea of their size. I know it's big; v
Near the NSA? (Score:2)
Parts of his dissertation where "removed".
He showed the choke points and critical links.
This plot of the large server areas would be fun to map and then visit.
Spot the NSA tap points
Work it backwards (Score:2)
What are the current server specifications necessary to support 50K virtual instances of an O/S? How many boxes does that need to map down to in order to maintain sufficient redundancy and power efficiency? Now, how big a datacentre do you need to support that number of servers and hence what space/power/air conditioning requirements would you actually need?
Google's container-based datacentre is about 1300 servers. If each of those is capable of running multiple instances, that would bump up the total.
OVH (Score:2)
they had 40K at the last official count and their new datacenter has a 50K capacity and filling quick (+3 bays/day 7/7). Not surprising given they offer the cheapest dedicated one can find ($15/mnth no contract: Atom 1.6Ghz, 512MB ram/2GB flash for swap, 10GB iSCSI disk, unlimited bandwidth).
1&1 (Score:2)
Never head of this company.
Re: (Score:1)
a huge hosting company from Germany, most customers are in Europe (but afaik they sell in the US, too).
you sub-6-digits old-timers should try to accept the i18n of /. :P
Re: (Score:3, Informative)
sorry for the second reply but according to the homepage ( http://www.1und1.com/ [1und1.com] [in German]) the company has over 7.2 million customers in Germany, UK, France, Spain, Austria and the USA running on more than 55k servers.
I'm selling my intelligence... (Score:1)
Re: (Score:1)
Wow, that guy must think himself a lot smarter than he actually is, where did you find the link?
Physical, obviously (Score:1)
MS Cloud is growing very quickly (Score:2)
While I can't speak for the others, I know Microsoft are growing their online infrastructure hugely now; at approx rate of "one facebook worth of servers every month" as one guy told me.
Most of it's consumer Live stuff, but there are plans to expand corporate services too so I understand.
Re: (Score:1)
Your average server rack is only 42U in height...50 discrete systems is extremely high density. Even with blades you rarely get past about 64 per rack...Most datacenters have a mix of machine sizes so the average density quickly drops below 40. This also drops again once you put in all the switching and routing necessary to support these networks.
I work for a VAR in our integration facility and the highest densities we ever see are 64 BL460c HP blades in a 42U rack.
Not Web Servers, but ... (Score:3, Funny)
I imagine there is a fairly large server farm somewhere near Ft. Meade, MD.
Blizzard? (Score:1)
All those WoW and Diablo fanboys have to go somewhere....
new industry dedicated to this (Score:2)
Modular expansion is fad. Server companies now recycle shipping containers as a row of racks with built in service, power distrib
"energy refineries" of the 21st century (Score:2)
I shamelessly stole this idea from Peter Huber of the Manhattan Institutes recent book The Bottomless Well. The trend of human commerce over the past couple centuries was to use more energy in more refined ways: wood, coals, whale-oil, petroleum, electricity, solar
Google has an order of magnitude more (fact) (Score:2)
I had a phone interview with google around '05, I think, and was *told* by the interviewer that they had over half a million (physical) servers. I don't remember the exact number, but it was between 500k and 600k servers.
mark
GSI Commerce could have a bit. (Score:1)
I know the company I work for GSI Commerce has one of the largest networks in the US, behind (if I recall correctly) Google, Amazon, Wal-Mart and Ebay. Now networks are not servers, but they have well over 100 partners that they represent with many's websites hosted by GSI. But still probably not anywhere near 50,000.
and wikipedia... (Score:1)