[c-nsp] Customer Facing Interfaces: Policing vs. Shaping

[Boyan Krosnov]

In a previous life I was helping run an ISP in the strange (for some)
environment of Bulgaria.
Bandwidth that the customers could afford was often not enough for their
needs. Customers were often trying to pull much more than their limit.
I'm talking about dedicated connection customers here, not small
residential ones. The bandwidths that were usual for one of these
customers at this time were between 128kbps and 4Mbps. It was quite
common for customers to be at 100% their capacity more than 50% of the
time.
International links were quite expensive. Capacity _to_ Western Europe
was much more expensive than the transit you would buy there. There were
only a couple of companies selling international capacity in the
country, only one of them using terrestrial links. All the other players
were buying transit from these two companies. No other company could
afford synchronous terrestrial transport to Western Europe. The
landscape has changed for good after I left the country :)

With CAR attached on output on a customer interface I have personally
seen _average_ drop rates above 5% of all packets. Peak drop rates of
more than 20% were not unusual. Some customers (especially ISPs) were
complaining about high drop rates. Some customers on shapers, were
complaining about high delays. For CARs we were using the X 2X/8 4X/8
rule and even gone to larger buckets in specific "high pressure" cases.
We were using CARs for most customers because of performance issues we
had with shapers, on the overutilized (remember it's a low-cost
environment) cisco routers. For individual customers we were forced to
switch to traffic shaping in order to meet their expectations.

My personal observation is that shaping provides lower drop-rates
compared to policing, regardless how well the policing buckets are
tuned. With shaping we could choose a delay/loss trade-off best suiting
the customer by tuning the queues. With policing we could not. On more
recent hardware/IOS it became possible to attach WFQ, instead of fifo to
the shaper, which gives you even lower loss and better fairness between
flows.

My vote goes to shaping not because it is the prevalent technology I
have used in the past, but because of my observations on how "well"
policing works under extreme conditions.

Cheers,
Boyan Krosnov
just another techie speaking for himself