Measurement-basedlLink dimensioning for the future internet

Abstract

In the future Internet network operators will control the physical networking infrastructure, but end users will be directly connected to companies, the Internet big players, that control essential online services and retain users’ content. Virtual networks will enable transparent and seamless connection between users and big players. The coexistence of many virtual networks on top of a single physical infrastructure will demand sophisticated approaches to fairly share and allocate link resources. Efficient link dimensioning approaches can certainly make the difference by: supporting operators on the optimal usage of their resources; ensuring that QoS metrics agreed with the big players are met; and providing end users with good levels of QoE. Focusing on proper allocation of link resources, we developed and validated link dimensioning approaches that are easy-to-deploy and accurate. Our starting point is an already validated dimensioning formula that requires traffic statistics that can be calculated from packet captures. However, packet captures are costly and often demand dedicate hardware/software. Our approaches estimate these statistics from coarser measured data, namely sampled packets and flows. Most network devices nowadays are supplied with technologies to perform these measurements, such as sFlow, NetFlow/IPFIX and OpenFlow. Our contributions can be divided in three major parts. First, based on inherited assumption of Gaussian traffic from the dimensioning formula, we showed that current traffic is Gaussian even though its characteristics have changed due to the recent advent of new online services (e.g., social networking, online storage and video streaming). This proved that the dimensioning formula used by us is still applicable to current Internet traffic. Second, we developed approaches to estimate traffic statistics for the dimensioning formula from sampled packets (e.g., obtained via sFlow), and from flows (e.g., from NetFlow). These approaches overcome the problem of coarser data and estimate statistics that ultimately yield accurate estimations even at millisecond timescales. Third, we introduced an OpenFlow-based approach to retrieve per-flow data from OpenFlow switches, and assessed the quality of this measured data both in a physical and virtual OpenFlow setups. We showed that per-flow measured data by OpenFlow is not yet suitable for link dimensioning.