Enterprise networks: Modern techniques for analysis, measurement and performance improvement

As the Internet evolves over the years, a large number of applications emerge, ranging from HTTP, FTP to P2P, and more recently multimedia streaming, on-line game and social networks, with varying service requirements in terms of bandwidth, delay, loss rate and so on. Still, the Internet traffic exhibits a high variability property - the majority of the flows are of small sizes while a small percentage of very long flows contribute to a large portion of the traffic volume. Several studies reveal that small flows are in general related to interactive applications - such examples consist of Web browsing, mail checking, DNS queries, and more recently tweets, posts, chats, etc. - for which one expects to obtain good user perceived performance, most often in terms of short response time. However, the classical FIFO/drop-tail scheme deployed in today's routers/switches

is well known to bias against short flows over long ones.

To tackle this issue over a best-effort network, a great deal of size-based scheduling solutions have been proposed in the last decade. The key idea is to favor short flows at the expense of long ones. Although appealing by offering small response time to short flows, most of them feature one or more significant drawbacks/limits: starvation to long flows, scheduling decision based on a single dimension, namely flow size, global modification of end hosts, and the overhead of flow state maintenance. In this thesis, we have proposed a novel and simple scheduling algorithm named EFD (Early Flow Discard), which is able to overcome all the drawbacks aforementioned.

In this manuscript, we first evaluate the performance of EFD in a single-bottleneck wired network through extensive simulations. We then discuss the possible variants of EFD and EFD's adaptations to 802.11 WLANs - mainly refer to EFDACK and PEFD, which keep track of the volumes exchanged in both directions or simply count packets in a single direction, aiming at improving the flow level fairness and interactivity in WLANs. Finally, we devote ourselves to profiling enterprise traffic, and further devise two workload models - one that takes into account the enterprise topological structure and the other that incorporates the impact of the applications on top of TCP - to help to evaluate and compare the performance of scheduling policies in typical enterprise networks.