Speaker: Daniel A. M. Villela
Title: Performance Analysis of Server Sharing Collectives for Content Distribution
To maximize revenue and user satisfaction, a content provider must provision its server resources to accommodate peak demands for its content. However, such provisioning is costly. As an alternative, we consider a novel framework in which content providers form collectives to meet peak demands without explicitly provisioning their individual resources for peak demand. Within these collectives, providers pool together their server resources, sharing the burden of servicing the individual demands of each collective member's content. Unlike previously considered resource-sharing paradigms, the preeminent goal of each provider in the collective remains the handling of the demand for its own content. Using analysis and simulation on fundamental queueing models, we compare the performance of a provider within a collective to its performance when operating in isolation. First, we consider a ``homogeneous'' collective composed of providers with identically provisioned resources and identical individual request loads. We verify that the blocking probability for each provider's content is lowered by several orders of magnitude by participating in the collective. We then examine ``heterogeneous'' collectives composed of providers whose resources and request loads differ. Surprisingly, we find that even for small differences, there can exist providers who are better off (achieving lower blocking probabilities) operating in isolation. To increase participation in the collective in these heterogeneous environments, we propose and evaluate novel thresholding techniques that limit the portion of each providers' resources that can be used by the collective. Our results show that thresholding significantly extends the set of heterogeneous configurations in which all providers benefit by participating in the collective.
Work conducted with Prof. Rubenstein