What you see before you is the Google Server, the commodity hardware Google uses to run all its operations. The machine is a standard Intel/AMD 2U server but each one has its own 12-volt battery, ensuring that the servers will maintain power in emergency situations, thereby alleviating the need for a large, nasty UPS.
They keep all of these servers inside shipping containers – they can fit about 1,160 in one container, and treat them like big fat black boxes.
What’s most compelling about this model is the banality of the hardware. Any one of us could create a similar server but to chain them together inside enclosed containers is an amazing feat. I can barely get my MacBook to work with my TiVo, let alone connect over one thousand servers to the world.
It’s only a harmless little piece of SKYNET. How cute.
am I missing something here or is this simply a google made blade server with a battery on it? It doesn’t seem too OMG to me. Also those large UPSes do more than just keep the servers running if the power goes down.
It’s the lovable simplicity of clusters… People often don’t know that they can be made out of “almost” standard off-the-shelf looking hardware… But wait a minute, Google did reported that they do use off-the-shelf components to build their cluster nodes and not server specialized ones.
Now multiply it by some hundreds and then a thousand and then you can begin to imagine the size of CADIE.
BTW, those batteries will surely run the node if an electrical event presents, but i would tend to think that they are there more like a way to seamlessly change from one power source to the other, rather than working merely as a UPS.
The local American Embassy auctioned some years ago a small trinket (size of a two door refrigerator) that could hold an array of car batteries to provide such power transition (APS i believe they called them), i take it that it might be easier to have it on the node instead.
does it strike anyone funny that this storey broke 1 April?
Assuming the servers run dual Xeon 2.8s (if each FLOP takes 3 instructions, thats around 933 MFLOPS per processor in ideal conditions) you’d be looking at 2.1 teraflops for the entire cluster, with 250kW peak power, about 8.6 MFLOPS per watt consumed.
The configuration to run those servers must be horrendous. Though, stick nginx on an outward facing server to act as a proxy to each of the other servers and it could be possible with a /21 subnet to the servers inside… maybe, just maybe.
I heard a Google engineer talk about this design at USENIX four or five years ago. Nice to finally see some pictures.