Resource usage prediction for application-layer multicast networks

We propose a memory and CPU usage prediction model for application-layer multicast networks. The predicted values are used in the distribution of end-hosts to overlay-hosts. Such a distribution enables us to limit the maximum resource consumption for a given node, leading to an efficient utilization of overlay-hosts and, finally, to an increased overall performance of the system. The model parameters are determined using training sets gathered from measuring resource consumption while distributing end-hosts to overlay-hosts. Both end-hosts and overlay-hosts are running on PlanetLab nodes, a testbed that provides researchers a real environment where nodes can become unreachable, network bandwidth can fluctuate and node processing capabilities can drop dramatically. Using the determined parameters, we show that our proposal can be used to estimate memory and CPU usage and to efficiently distribute end-hosts to overlay-hosts.