Just to discuss a few points, I don't believe are clear ([edit] or sufficient emphasized) in the TechNote.
When a shaper queues the "delayed" packets (unlike dropping them as a policer would), these queues can overflow and still result in dropped packets. ([edit] This is documented in the TechNote, along with the possible packet delay caused by such queuing. Both easy to overlook.)
It's also not well documented what kind of queue or queue the shaper uses. From experience, many router shapers seem to use WFQ for their delayed packets. ([edit] If WFQ or FQ is used, original packet sequence might be changed [not per flow, though] - policer doesn't reorder any packets.)
Both policers and shapers measure bandwidth utilization across a certain time period and enforce an average utilization, not an instantaneous bandwidth utilization. For instance if you police or shape a 100 Mbps link for 10 Mbps, neither behaves exactly like an actual 10 Mbps link would. ([edit] Tuning of the Bc and Be adjusts behavior.)
