On characterizing affinity and its impact on network performance

Abstract

An important component of simulation-based network research is the selection of nodes to a member group, such as receivers in a multicast group or web clients in a content delivery network. In a seminal paper, Philips et al. introduce an algorithm for generating member groups with different degrees of affinity (clusteredness) and show that affinity can have a significant effect on multicast efficiency. Subsequent studies applying this algorithm have all used the algorithm's input parameter as a method for classifying and comparing affinity groups. In this paper, we propose several distance- and expansion-based analysis metrics and find them to be better measurements of the true affinity of member groups. In three separate case studies (multicast, replica placement, and sensor networks), we demonstrate the benefit of classifying member groups by their true affinity in order to predict network performance variation. By systematizing techniques for measuring affinity, we open the door for more realistic and reproducible research in studies employing affinity-based member selection techniques.