Happy Packets
Randy Bush / 2003.09.30
As routing researchers, we frequently hear comments such as
o internet routing is fragile, collapsing, ...,
o bgp is broken or is not working well,
o yesterday was a bad routing day on the internet,
o change X to protocol Y will improve routing,
o etc.
And we often measure routing dynamics and say that some measurement
is better or worse than another.
But what is 'good' routing? How can we say one measurement shows
routing is better than another unless we have metrics for routing
quality? We often work on the assumption that number of prefixes,
speed or completeness of convergence, etc. are measures of routing
quality. But are these real measures of quality?
Perhaps because I am an operator I think the measure which which
counts is whether the customers' packets reach their intended
destinations. If the customers' packets are happy, the routing
system (and other components) are doing their job.
Therefore, I contend that, for the most part, we should be judging
control plane quality by measuring the data plane. And we have
well defined metrics for the data plane, delay, drop, jitter,
reordering, etc. And we have tools with which to measure them.
It is not clear that happy packets require routing convergence as
we speak of it today. If there is better routing information near
the destination than at the source, maybe there is sufficient
information near the source to get the packets to the better
informed space. This is not that unlike routing proposals, such as
Nimrod, where more detail is hidden the further you get from the
announcer.
If the routing system is noisy, i.e. there is is lot of routing
traffic, that may not really be a bad thing. We know convergence
time can be reduced if announcement throttling (MRAI) is lessened.
As long as network growth increases load on the routers below
Moore's law, it is not clear we are in danger.
So, happy packets to you.