ABSTRACT

This thesis presents a supervisory mechanism that is able to stabilize
any  non-acyclic,  stochastic,   multiclass  queuing  network  with  a
distributed scheduling  policy. The Traffic Intensity Condition has to
hold  as well  as some  mild assumptions  on the  service and  arrival
distributions.

This  mechanism is  based on  the Active  Idleness concept.  By Active
Idleness it is meant  the scheduler's  ability to force  inactivity on
the  server,  even  in the  presence  of  work. Although  this ability
appears  as  leading  to  a  waste  of  the  available  resources, the
important factor  is to look at  the queuing network  as a whole. What
might  appear  as a loss  of resources when  looking  to an individual
server, could in fact be highly  beneficial for the performance of the
entire queuing network.

The thesis presents one possible implementation of the concept, termed
Time  Window  Controller, and  illustrates  its  ability to,  not only
stabilize,  but  also  improve  the  performance  of some  stable non-
-acyclic, stochastic,  multiclass queuing  networks,  independently of
the performance measure under consideration.


KEYWORDS: Queuing  Networks,  Distributed  Scheduling,
          Stability, Idling Policies, Active Idleness.