Recent events such as the Mediterranean cable cuts of December 2008 remind us that although this ability to self-repair is to some extent true, in reality large sections of the net are actually connected to the other parts of the world by a small number of links, rather than a widespread grid of connections.
These links are vulnerable, and just a few breaks can seriously impact connectivity in these regions. In his article It’s 2008 – Do You Know Where Your Internet Cables Are? (concerning a similar event earlier in the year) Alistair Croll explains the problem of having two key cables very close together, so both can easily be damaged by the same cause, whether because of a malicious act or by accident.

So while the internet is able to recover in a reasonable period of time after damage to the routers, it is more susceptible to the cutting of the links between the routers. Satellite links can take up some of the slack, but bandwidth is limited.

Thus the idea that the internet is resilient to damage is at least partially a myth. This concept is explored in Neal Stephenson’s novel Cryptonomicon, where the connectivity of undersea cables carrying internet traffic plays a key role in the plot.

As Croll says

The fact that these two cables — which are supposed to provide redundancy — were so close together should be cause for alarm. When companies build data centers, they go to great lengths to ensure that power supply into the data center comes from multiple directions in order to avoid interruptions. But geographic and technical constraints mean some places are particularly crowded, much like modern versions of the Suez and Panama canals. And with this concentration comes vulnerability.

While we may be able to make numerous improvements in the underlying internet protocols, the reality is that the internet is vulnerable to physical problems. There are some enhancements that could be made to the net that would alleviate these problems to some extent; more distributed caching of content around the world would mean that sites, or more importantly media, would still be accessible even though the ‘real’ site is unreachable. A ‘grid’ approach to media delivery is becoming popular, as content providers are reducing the demands on their own servers (and their links) by getting consumers to share the burden of transmitting files to other consumers, in a peer-to-peer system. These grids were not designed to make the internet as a whole more resilient, but it doesn’t take an enormous leap of imagination to see how an extension of the idea could have this effect.

Ultimately, the best way to get internet resilience is link redundancy. Lots of cables, lots of different links, all going via different physical routes, creating a more mesh-like network which can tolerate breaks.