Abstract:

Differentiated Services (or just DiffServ) is a very scalable
architecture for delivering QoS in the Internet.  This scalability is
achieved by maintaining a per-class (and not a per-flow) state at the
core routers. However, this exact property also makes ensuring
end-to-end QoS very difficult, since one has to verify that the local
resources allocated to each class are sufficient to support the
requirements of all the flows in that class. This is particularly
challenging when considering accumulative requirements such as delay or
loss.  This talk addresses two aspects of this problem.

First we address the translation of the end-to-end delay requirements of
flows to a specific resource allocation at the different routers. The
required end-to-end QoS level is usually specified in a Service Level
Agreement (SLA). The problem is thus to find a way that maximizes the
number of flows for which the delay specified in the SLA is satisfied in
congested conditions. The work evaluates the importance of both local
algorithms that guarantee a constant ratio between local waiting times
of packets belonging to different service classes, and global probing
techniques that allow classifying the packets into the different service
classes according to their SLA defined delay, in maximizing the number
of flows for which the delay constraint is satisfied.

Then we concentrate on the probing mechanism, trying to minimize the
amount of overhead needed. We propose a new probing technique called,
Adaptive Probing, in which the routers along the flow path are
responsible to discover and notify a centralized entity when a violation
of the SLA occurs (or soon to occur).   We study the performance of this
new algorithm through theoretical analysis and extensive simulations.
Our results indicate that in addition to dramatically reducing the load
from the centralized entity, the amount of network traffic needed is
relatively small, and this new monitoring scheme scales well both with
the  number of hops and the network load.
                                                                                
This is a joint work with Alex Dvorkin, Constantine Elster, and Ran
Wolff.