Inproceedings,
A Cellular Neural Networks Based DiffServ Switch for Satellite Communication Systems
Proceedings of the First International Conference on Advanced Satellite Mobile Systems ASMS 2003, (July 2003)
Abstract
Recent developments of Internet services and advanced compression
methods has revived interest on IP based multimedia satellite communication
systems. However a main problem arising here is to guarantee specific
Quality of Service (QoS) constraints in order to have good performance
for each traffic class. Among various QoS approach used in Internet,
recently the DiffServ technique has became the most promising solution,
mainly for its simplicity with respect to different alternatives.
Moreover, in satellite communication systems, DiffServ policy computational
capabilities are placed at the edge points (end-to-end philosophy);
this is very important for a network constituted by one satellite
link because it allows to reduce the implementation complexity of
the satellite on-board equipments. The satellite switch under consideration
makes use of the Multiple Input Queuing approach. Packets arrived
at a switch input are stored in a shared buffer but they are logically
ordered in individual queues, one for each possible output link.
According to the DiffServ policy, within a same logical queue, packets
are reordered in individual sub-queues according to the priority.
A suitable implementation of the DiffServ policy based on a Cellular
Neural Network (CNN) is proposed in the paper in order to achieve
QoS requirements. The CNNs are a set of linear and nonlinear circuits
connected among them that allow parallel and asynchronous computation.
CNNs are a class of neural networks similar to Hopfield Neural Networks
(HNN), but more flexible and suitable for solving the output contention
problem, inherent of switching systems, for VLSI implementation.
In this paper a CNN has been designed in order to maximize a cost
functional, related to the on-board switch throughput and QoS constraints.
The initial state for each neural cell is obtained looking at the
presence of at least one packet from a certain input logical queue
to a specific output line. The input value for each neural cell is
a function of priority and length of each input logical queue. The
versatility of neural network make feasible to take the best decision
for the packet to be delivered to each output satellite beam, in
order to meet specific QoS constraints. Numerical results for CNN
approach highlights that Neural network convergence within a time
slot is guaranteed, and an optimal, or at least near-optimal, solution
in terms of cost function is achieved. The proposed system is based
on the IETF (Internet Engineering Task Force) recommendations; this
means that traffic entering the switching fabric could be marked
as Expedited Forward (EF) or Assured Forward (AF), otherwise handled
as Best Effort (BE). Two Assured Forward classes with different emission
priority have been implemented, taking into account time spent inside
the logical queue and its length. Expedited Forward traffic is typical
of services to be delivered with the maximum priority, as streaming
or interactive services. The packets, belonging to services that
need a certain level of priority with low packet loss, are marked
as Assured Forward. Best Effort traffic is related to e-mail or file
transfer, or other that have not particular QoS requirements. The
CNN used to solve conflict situations act as an arbiter for all the
output links. Differently from other Multiple Input Queuing approach,
where one arbiter for each output line is present, in proposed approach
there exist only one arbiter that make the best decision. The selected
rule has been defined in order to give priority to packets, according
to opportunely defined functionals characteristic of each traffic
class, under the constraint that no more than one packet can be delivered
to the same output line. The functionals depend on queue length and
time spent inside the queue by front packet. The performance of the
proposed DiffServ switch has been derived in terms of delay and jitter;
buffer occupancy has been analyzed for different configuration, such
as a unique common buffer, one buffer for each input line, one buffer
for each input line and each priority class. The obtained results
highlight an high flexibility of satellite switch with CNN, taking
into account that functional used to calculate priority of each queue
could be easily changed, without any complexity gain nor change in
CNN structure, in order to consider different traffic characteristic.
Numerical results show that proposed algorithm outperform the switches
based on Multiple Input Queuing, that use strictly priority methods,
in terms of delay and jitter. Different buffer size have been also
considered in order to analyze packet loss for CNN switch algorithm,
comparing different configuration described above. The good behavior
of the proposed DiffServ switch has been verified in the case of
traffic with pareto distribution for packet length and a geometrical
distribution for packet interarrival time, highlighting good performance
in terms of delay and jitter. Numerical results also demonstrate
the stability of this method for heavy load traffic; in particular
maximum permitted load is higher for higher priority classes.
