trip-time and packet loss ratio has been sug-
gested. The delay and packet loss of a flow is
greatly influenced by the number of flows in
progress using the same set of resources. This
means that the packet scale performance of a
given flow is strictly determined by flow level
dynamics (i.e. number of flows and their charac-
teristics). A simple model of this is quoted in
[Robe01], referring to the processing sharing
analogy when looking at a single resource.
Although still being a simplified model, two
central observations are made: i) the perfor-
mance depends primarily on expected traffic
demands and only marginally on parameters
describing distribution of file lengths; and ii)
performance tends to be excellent as long as
expected demand is less than available capacity.
The latter proposes that service differentiation
can only be effectively obtained for a limited
area of the workload when there are several ser-
vice classes to be handled with their correspond-
ing requirements.
For streaming traffic, the algorithms in TCP may
not be in effect, e.g. as UDP could be applied.
This means that the characteristics would be
more determined by the inherent nature of the
traffic source (as an open-loop control would be
present). Thus, it is simpler for a source/applica-
tion to specify the required transfer service,
which for instance can be input to an admission
control function (as well as resource reservation
and others).
Integrating elastic and streaming traffic on the
same resource units may allow for increased
efficiency. By giving priority of the streaming
flows, they could experience a resource that is
(almost) loaded as if they were the only active
flows. Then, elastic flows could be served when-
ever the resource is not used by the streaming
flows. However, this may introduce long delays
for the elastic flows during some periods. An
approach is to restrict the load from the stream-
ing flows, ensuring that some capacity is avail-
able for the elastic flows. This is an argument
for introducing admission control that operates
also on streaming flows.
When the capacity of the resource is limited, it is
generally assumed that some form of admission
control has to be present to ensure that active
streaming flows receive the delay and packet
loss requirements they demand. To avoid keep-
ing a detailed list per flow, a measurement-based
approach could be used, operating on the aggre-
gated flow. It is argued in [Robe01] that such an
aggregated measure might not be very precise,
as the elastic traffic flows would tolerate some
variation in their available service rate.
In some cases it is argued that so-called over-
provisioning may make the need for traffic han-
dling mechanisms obsolete, including admission
control. Apart from the economic argument, it
might also be difficult to provide the amount of
capacity on certain portions, like on the access
line if no technical solution is available. Another
argument is the request for service differentia-
tion. As quite a few models show, a “pure” Diff-
Serv model may have a narrow scope for effi-
cient differentiation that is close to an overload
situation. Hence, other means for differentiation
would be asked for. Therefore, providing differ-
entiated admission criteria is one group of means
that could be introduced.2.4.3 Admission Control Taxonomy
A number of issues for describing an admission
control mechanism are described in this section.
The issues are not independent as some combi-
nations may be preferred or even needed for the
admission control procedure to function.- Dynamic versus static. A dynamic mechanism
is able to adapt to the changing situation, for
example captured by measuring traffic loads.
It is obvious that measurements for better
assessing the link utilisation and characteris-
tics of the traffic flows will lead to improved
throughput. On the other hand, running con-
tinuous measurements may be fairly demand-
ing on the routers. Hence, finding adequate
measurement arrangements is a central chal-
lenge. - Explicit versus implicit. When explicit control
is used, related information is exchanged be-
tween the end system and the network. That
is, protocol elements are defined which ex-
press the request for resources and the grant-
ing/rejection of resources. In case implicit
control is applied, there will be no information
exchange before the information transfer. An
example is when the source simply starts to
send packets and the network decides whether
or not these packets can be forwarded without
informing the source of the decision. Thus, the
source has to apply other means for finding
out if the transfer was successful or not (e.g.
time-out on acknowledgements). - Scope and objective function applied. When
deciding whether or not to accept a request,
different scopes and different sets of variables
can be implemented. This is further described
below. - Traffic flow aggregates and characteristics.
The admission control may work on a range of
traffic flow aggregates and use various means
for describing the traffic flows. Examples of
bit rate measures are peak rate, mean rate and