Improving the Reliability of Internet Paths with One-hope Source Routing

Motivation: Scalable One-hop Source Routing proposes a simple, scalable
approach
to recover from Internet path failures.

Key Points:

1) The paper conducts a broad measurement study of Internet path failures
on a collection of 3000 Internet nodes. It observers that when failures do
occur, many appear too close to the destination to be mitigated through
alternative routing techniques of any kind.

2) Failure has been defined as a sequence of packet losses that would have
a significant or noticeable application impact. A loss incident was elevated
to a failure if anf only if the loss inceident began with 3 consecutive
probe
 losses and the initial traceroute failed.

3) Last hop failures are found to be pre-dominant in case of braodband hosts
(thus low benefit from SOSR). However, for popular servers the last hop
failures
were rare, thus a technique like SOSR would be likely give good results.

4) A small number of paths experience a very large number of failures. Most
paths experience a small but non-zero number of failures and experience less
than 15 minutes of downtime over the week.

5) Concept of One-hop source routing - After a node detects a path failure,
it selects one or more intermediaries and attempts to reroute its packets
through them. If the resulting indirect path is sufficiently disjoint from
the default route the packets will flow around the faulty components and
successfully arrive at the destination.

6) By measurements it was established that a random-4 policy, where the node
 picks up 4 nodes randomly to route through them makes a reasonable tradeoff
between effort and proabability of success. It can be improvised, by
choosing
k-1 nodes randomly and the last successful node from history.

7) SOSR is successful at routing around non-last hop network failures and
it has very little overhead. SOSR is likely to achieve close to the maximum
 achievable benefit of alternative routing without the need for path
monitoring,
history or a-priori knowledge of any kind.
Received on Tue Oct 17 2006 - 05:36:19 EDT