Jeffery P. Hansen
Research Associate Professor
Curricula Vita (ps, pdf)
While network bandwidth reservation protocols such as RSVP can be used to provide consistent service levels, the bursty nature of multimedia traffic results in significant underutilization of resources due to the worst-case assumptions of these approaches. Since most distributed multimedia applications can tolerate a small amount of packet loss without significant degradation in delivered service, Dr. Hansen believes that hard reservations are often unnecessary, and that probabilistic guarantees yield a more effective tradeoff between quality of service and resource utilization.
Probabilistic guarantees for network bandwidth are enforced through a combination of admission control algorithms and mechanisms, collectively called RPM (Resource Priority Multiplexing), developed by Dr. Hansen. In the RPM paradigm, the bursty nature of multimedia flows are modeled by Markov modulated Poisson processes. Each flow is also assigned a guarantee level representing a lower bound for the expected value on the fraction of packets from that flow for which delivery is guaranteed. Guarantee levels can range from 0% for best-effort delivery to 100% for hard guarantees. The basic idea of RPM is to maintain a set of time multiplexed priority modes which determine the order in which packets will be dropped when there is a resource shortfall. The relative holding times for these modes are computed by solving a linear program formulated from the per-flow traffic models and guarantee requirements.
Among the challenges in resource management for QoS is allocating resources such that user satisfaction, as measured by per-task ``utility functions'', is maximized while QoS reconfiguration of admitted tasks is minimized. QoS reconfiguration occurs when the QoS set-point, a vector specifying quality values for each QoS dimension (e.g., a video flow with cif resolution at 30 frames-per-second), for an admitted task must be downgraded to release resources to satisfy a new task arrival. This can occur when admission control policies fail to anticipate future task arrivals and over-commit resources.
In dynamic QoS models proposed by Dr. Hansen, future demand is modeled as a pseudo-task deriving utility for maintaining a resource reserve. The utility function for reserve resources is such that when resources are scarce, the marginal utility for keeping resources in reserve is high, while when resources are plentiful, the marginal utility is lower. QoS set-points are assigned to incoming tasks by maximizing the joint utility between the reserve resource pseudo-task and the incoming task. He has shown that this admission policy can result in global utility (the sum of the utility over all tasks) which is almost as high the theoretical maximum with a negligible rate of reconfiguration events.
Dr. Hansen has also worked in developing tools for scientific visualization and has additional interests including data mining, computer security, formal verification of computer systems, and logic synthesis.