We’ve got some new blade frames here that pack something like 26 blades into a rack, and each blade has four dual-core CPUs. That’s 104 processors in four square feet of floor space.
Assuming you have the electrical power to feed it, and HVAC to cool it, half a million servers (assuming 1 CPU = 1 server) would fit in five thousand racks or 20,000 square feet. Add space for aisles, and it’s still probably under 40,000 square feet. Mind you, this is with everything absolutely crammed together. A big building, certainly, but not unthinkably large.
I was curious for a size comparison, and a quick Google later, found that a typical Costco store is in the neighborhood of 175,000 - 200,000 square feet.
Just watch out that the rack doesn’t melt from all that heat in one spot.
Electrical power becomes challenging at that sort of density as well - 42 servers in one rack is going to consume in the area of 75-100 amps of 120-volt power and kick out a similar amount - 8000 - 10,000 watts of heat.
Not that you can’t do it, but you really need to have a firm grasp of power distribution and air conditioning to make it work.
Indeed. At my former client, we had our racks hosted at Peer 1’s colo center in Manhattan. Our densest one had about 30 machines crammed into it (most were 1U but some 2U) as well as two GigE switches, a KVM station, a 3U NAS box, four remote-rebooter devices (mounted on the back) and other miscellaneous stuff. It put out a lot of heat.
Fortunately, Peer1 has two gigantic HVAC units that produce a constant 10mph wind in the datacenter. It’s also so loud that it’s difficult to talk in there (or hear yourself think.)
Personally, I prefer programming from the comfort of an office chair. Datacenter work is the pits.
75-100 amps times 120 volts is 9000-12,000 watts, and all of that is going to end up as heat sooner or later (mostly sooner). I’m not sure where your figure of 8000 - 10,000 comes from.
We had a Google talk at university last year. They’ve got their own specially crafted database software designed for handling the gargantuan levels of data and specially designed query languages for querying the database. Commercial databases don’t cut it, apparently.
No, they wouldn’t. Commercial databases are based on being generically usefull, and optimizable if you know exactly everything there is to know about how the database does it’s thing. Most people don’t, and all the reference material is very arcane.
Just the fact that you have to do everything over sockets already kills you. Not to mention that all data is transmitted as text, so you need to convert all numeric data back and forth, etc.
The major thing that commercial databases solve for you is that “it’s already made” and can be called from anywhere.
