Google: turning cash into cache

Google and dark fiber

Interesting theory on why Google is buying up miles and miles of ‘dark fiber’ hi-speed optical network (they have been doing this since at least 2002).

The probable answer lies in one of Google’s underground parking garages in Mountain View. There, in a secret area off-limits even to regular GoogleFolk, is a shipping container. But it isn’t just any shipping container. This shipping container is a prototype data center. Google hired a pair of very bright industrial designers to figure out how to cram the greatest number of CPUs, the most storage, memory and power support into a 20- or 40-foot box. We’re talking about 5000 Opteron processors and 3.5 petabytes of disk storage that can be dropped-off overnight by a tractor-trailer rig. The idea is to plant one of these puppies anywhere Google owns access to fiber, basically turning the entire Internet into a giant processing and storage grid.
While Google could put these containers anywhere, it makes the most sense to place them at Internet peering points, of which there are about 300 worldwide.

Two years ago Google had one data center. Today they are reported to have 64. Two years from now, they will have 300-plus. The advantage to having so many data centers goes beyond simple redundancy and fault tolerance. They get Google closer to users, reducing latency. They offer inter-datacenter communication and load-balancing using that no-longer-dark fiber Google owns. But most especially, they offer super-high bandwidth connections at all peering ISPs at little or no incremental cost to Google.
from “Google-Mart” (Cringely)

King of cache

Having a datacenter close to any customer is of course the best way to do swift and reliable caching. Presumably not for its primary activity web search, since caching responses for search makes not much sense (see Pareto doesn’t do search). That kind of local storage/network capacity can nevertheless be used to cache other stuff:

And all that processing power (with some simple extrapolation: 17 Teraflops per datacenter block) could be used to index, filter and convert data or just rent out for distributed computing.

Orders of magnitude

How big is a petabyte? A thousand terabyte, but that doesn’t say much. “If digitized with full formatting, the seventeen million books in the Library of Congress contain about 136 terabytes of information.” (Berkeley 2003) So a datacenter of 3.5 petabytes, that’s about 25 libraries of Congress. At the fastest network speed we know, 10 Gbit Ethernet or 108 TB/day, it would take 32 days to fill one of such datacenters. And it demands something between 5 and 10 Megawatt of power.


💬 Google