Bradford D. Boyle
Successful network monitoring often involves a collection of measurements at a large number of geographically disparate locations that need to be communicated back to a central location for processing to observe the state of the network. A classic example of this is estimating the sender-destination traffic intensity from repeated measurements of aggregate link loading. Reporting these measurements back to a central location artificially increases the load on the network. We are investigating recent advances in multi-terminal source coding to develop techniques that exploit the correlations amongst the collected measurements that are induced by the structure of the network. The goal is to minimize the network overhead , as a function of the source encoder rates, needed to convey observations to a central fusion center.An often encountered constraint with network monitoring is the inability to collect measurements at every network location. This limitation can be overcome to a degree by injecting probing traffic and measuring the outcomes of these probes at a subset of network locations. By utilizing different probing techniques (i.e. multicast, unicast striping, etc) it is possible to infer both the routing topology between the source and measurement sites. Additionally, one can characterize the performance of internal network components that cannot be directly measured. We are studying the correlations that are induced by the network structure in order to identify a subset of measurement sites that will yield the most information about the rest of the network.