Buffer management and quality of service predictability for best effort traffic

The increasing use of the Internet and computer networks in general to run old and new applications is putting a lot of pressure on how networks are designed and built to support those applications adequately. Networking technologies have been designed to offer different types of services, each one meant to address the main requirements of certain applications. Each type of service contains a particular set of algorithms and mechanisms that enforce those requirements within the network. However, the most successful service so far has been surprisingly the one that does not provide any kind of performance guarantees to applications. From the network and applications' point of view, this service, commonly known as the best effort service class, is the most simple to understand and implement. Applications do not need to know their traffic requirements and characteristics to ask the network for specific resources to satisfy them. On the other hand, networks only need a scheduling discipline and a buffer management scheme. However, one drawback of this service is that its quality depends on the number of applications joining and leaving the system at any given time, and as a result, applications can receive not only bad performance but also a very unpredictable service.; Buffer management strategies have been widely recognized as a critical component in all networking technologies. In best effort services, the Buffer Management Scheme (BMS) is perhaps the most important and influencing mechanism since the scheduling mechanism of choice is usually First Come First Served (FCFS). Buffer management schemes for best effort services were initially created to combat network congestion but nowadays they have also assumed other roles, such as the improvement of the application-level performance, the network utilization and the predictability of the service. This dissertation studies the problem of buffer management in the best effort service category of Asynchronous Transfer Mode (ATM) networks and proposes a novel buffer management algorithm called Preemptive Partial Packet Discard (pPPD) that exhibits important gains in performance over the existing ones, both in absolute and relative terms. Using simulations, we show how pPPD improves the application-level loss and delay performance metrics, the network utilization and the application-level predictability, allowing applications to predict the quality of the best effort service and adapt accordingly. In addition, pPPD has a very important and unique characteristic for a buffer management scheme: it is very flexible. pPPD can be easily changed to suit or satisfy particular needs. Besides the “plain” pPPD, five pPPD variants are also proposed. pPPD-hops and pPPD-weighted are proposed to increase the application performance over multi-hop environments; pPPD-clr and pPPD-time are proposed to modify the best effort service model to improve the service predictability for some applications or classes of applications supporting relative traffic differentiation; and pPPD-tcp is proposed to improve the performance of TCP traffic over ATM networks.; In addition, the applicability of the buffer management algorithms used in ATM networks is studied in general networks as the performance of the generalized versions of these schemes is analyzed using numerical analysis and simulations.