Admission Control versus Policy Control

At this point we should differentiate admission control and policy control:
 Admission control maintains information about the available resources from a
network entity.
 Policy control checks the administrative entitlement to the requested QoS
(which may be for an individual user or aggregated services).
We have said that offered traffic distribution and offered traffic properties determine
the provisioning of radio and network bandwidth. This includes delivery and memory
bandwidth. We are discussing packet-routed networks that are, essentially, queuing
networks. It is the process of queuing that delivers bandwidth efficiency by degrading
QoS for users who hopefully don’t notice or don’t care, in order to deliver better service
to those who do notice or do care and are willing to pay for the privilege of being
given priority access.
Queuing networks do not like bursty bandwidth. Bursty bandwidth fills up buffer
bandwidth. When buffer bandwidth overflows, packets are lost and, if using TCP,
packets have to be re-sent, absorbing additional delivery bandwidth, as well as RF
power. This points to a direct relationship between buffer bandwidth/memory bandwidth
and packet loss, which in turn has a direct relationship with end-to-end delay
and delay variability.
To quote from an IEEE paper (IEEE Communications, January 1999, Zheny, Atiquzzaman,
Kouz, Sahinoghu, Tekincy): “Under many conditions, it can be observed that a
linear increase in buffer size results in an exponential increase in packet loss.” A typical
performance for a 2-Mbyte buffer at 60 percent usage might be a 4 x 10-2 packet drop
rate. A64-Mbyte buffer reduces the drop rate to 3 x 10-3. It is also not only the size of the
buffer but the way the buffer memory is partitioned; a router memory that clumps data
in batches may actually increase rather than decrease burstiness.
Further, there is a philosophical issue when considering how bursty bandwidth
should be managed. You have a choice: You can either make networks applicationaware
(that is, the network determines the resource requirements of the application),
or you make applications network aware (that is, the application can validate, qualify,
and quantify the availability of current network resources, and possibly compare availability
and cost in other networks).
Needless to say, network operators prefer the first option—put the network in control.
Unfortunately, the second option works better, and it potentially confers more
benefits, including cost-saving benefits, to the user.