Deployment Considerations of Implementing Differen (decides) Concluded WG
Note: The data for concluded WGs is occasionally incorrect.
|WG||Name||Deployment Considerations of Implementing Differen|
|Area||Transport Area (tsv)|
|Dependencies||Document dependency graph (SVG)|
Charter for Working Group
The IETF WG on Differentiated Services (diff-serv) has been meeting now
over a year now and has made excellent progress. The framework,
DS Field definition, and format for traffic conditoners document, and
importantly , two PHB definitions have been specified.
So what is the problem?
The Integrated Services WG (int-serv) (with its related groups such as
ISSLL and RSVP and others), has also defined a set of protocols and
mechanisms for enhanced services in the Internet. The difference between
int-serv and diff-serv approaches is that int-serv (at least the
and Controlled Load services) is amenable to analysis. Effective
calcualations, and the thesis by Parekh shows how the admission tests
bucket characterised traffic and delay bound for the path can be
calculated. The difficulty with IS has been in designing
implementations that scale (hence a great deal of work in state
aggregation in RSVP, and in fast generalized port specification
classifiers, as well as the work on efficient queue data structures
and insert/retrieve algorithms, such as WF2Q and approximations such
as SFQ). Indeed, there are network QoS calculi (by Rene Cruz and also
by Jean-Yves le Boudec).
Meanwhile, with diff-serv, the difference is that there can be no
obvious analytic theory of diff-serv. A path can be constructed out of
a sequence of hops each with a PHB and some associated SLA. However,
the service (and provisioning of service) necessarily depend on the
actual network topology and traffic conditions that prevail.
This is a positive aspect of diff-serv, since it gives providers (and
router vendors) a lot of design freedom in how they deploy actual
services (and associated tarrif structures).
However, to evaluate a diff-serv PHB is now a complex task, and
requires simulation or measurement. To date, only modest simulations
and measurements have been carried out.
[Aside: of course the exact same argument
applies to figuring out call blocking
probabilities in int-serv type networks]