Spare a Little Change? Towards a 5-Nines Internet in 250 Lines of Code

From its beginnings as a single link between two research institutions in 1969, the Internet has grown in size and scope, to become a global internetwork connecting over 700 million computers, and 1.7 billion users. No longer a niche facility for scientific collaboration, the Internet now touches the lives of the world’s population, irrespective of their occupation or geography. It is used by people the world over, to pay bills, read the news, listen to music, watch videos, telephone or video-conference friends and family, and much more. The Internet is the premier communications network of our age.

Unfortunately, however, there are some respects in which the Internet lags the networks it replaces. In particular, with respect to reliability, the Internet falls far short of the Public Switched Telephone Network which proceeded it. Whereas the PSTN sought, and often delivered the vaunted “five nines” of reliability, the Internet struggles to compete. As for the cause of this reliability shortfall, available evidence indicates that much of the shortfall is due to the unreliability of IP routers themselves.

Given the importance of a reliable Internet to contemporary society, vendors and researchers have proposed a number of solutions to either improve the reliability of individual IP routers, or to make networks more resilient to the unavailability of a single router. While having some promise, these existing solutions face significant obstacles to widespread deployment. Thus, in this dissertation, we endeavor to find or construct a practical, readily deployable, method for mitigating the outages caused by IP routers.

To achieve our goal, we take inspiration from previous proposals, which advocated the use of link migration. These proposals improve network resilience, by moving links away from a failed (or failing) router, to an in-service router. To understand the constraints of a practical solution, and resolve the limitations of previous proposals, we conduct extensive experimentation, and study source code and protocol specifications. Using the insights produced by these studies, we construct a practical, readily deployable migration solution with sub-second outage times.