Collaborative Research: A Control Theoretic Approach to the Design of Internet Traffic Managers

University Of Arizona, Tucson AZ

Investigators

Abstract

The scalability of the Internet hinges on our ability to tame the unpredictability associated with its open architecture. This project investigates the development of basic control strategies for reducing traffic burstiness and improving network utilization. Such strategies can be applied through Traffic Managers (TMs)-special network elements strategically placed in the Internet (e.g., in front of clients/servers or at exchange/peering points between administrative domains). We believe that the incorporation of such control functionalities will be key to the ability of the network infrastructure to sustain its own growth and to nurture the Quality-of-Service (QoS) needs of emerging applications. Although there have been some recent advances in building network elements capable of wire-speed processing, there is a need for fundamental research into the basic QoS control capabilities that these TMs should implement. This set of capabilities have to be identified and implemented in a programmable, scalable architecture that allows for the easy and effective composition of services. Such a flexible architecture is highly desirable as the Internet continues to evolve and users demand new kinds of service for their applications. TMs should be capable of quickly inspecting and classifying packets as they go by (e.g., marking packets into precedence classes), and should control the transmission of these packets (e.g., through pacing, scheduling, or selective dropping) to ensure desirable properties (e.g., satisfaction of jitter requirements, compliance with TCP friendliness, or improved fairness across flows). In this proposal, we will address the design of dynamic dos control programmable TMs. We focus on basic capabilities that could be employed at different levels of the control architecture. These capabilities include differentiated, aggregate and proxy controls. The following are examples of how such control strategies would be employed by TMs. Differentiated Control enables TMs to route flow aggregates with divergent characteristics on separate communication paths. Unlike traditional routing, our routing metrics will respect bursitis measures, such as self-similarity and traffic correlation: Aggregate Control enables TMs to use congestion control mechanisms for collections of flows that share the same bottleneck. Unlike traditional congestion control, "Congestion-equivalent" flows are identified based on measures of relationship (such as cross-correlation and cross-covariance) and managed as a set; Proxy Control enables TMs to filter out variability (e.g., loss, delay jitter) at shorter time-scales. Such a functionality is crucial for improving the stability and effectiveness of control mechanisms that operate over longer time-scales (e.g., end-to-end). Unlike traditional a-hoc proxy approaches, our approach will take into account the length and characteristics of the control loops that get formed between the TM and the end-systems. Our design will be based on mathematical foundations from control theory and wavelet analysis. These methods enable thorough analysis and control of system dynamics at different time-scales and an understanding of the complex interactions among them. Specifically, functionality's at different levels of a TM architecture will be developed based integrated control-theoretic models. These models will account for "nested" control loops that are driven by system characteristics, which are identified using wavelet analysis of passive measurements. TMs that are designed in such and integrated fashion, could increase flow throughput, reduce flow jitter and response time, and improve the stability, utilization, and scalability of the network. We plan to implement our dos controls in a tested deployed in a controlled local setting as well as over the Internet. Our implementations will be based on emerging technologies, such as Diffserv and MPLS, and will be stressed by bandwidth-and QoS-demanding applications. Our testbed will provide a programming interface to softservices, in which capabilities can be turned on or off and control parameters can be dynamically adjusted. To this end, we have secured the support of industrial research laboratories and start-up companies-namely Lucent's Bell Labs, Cisco Systems, Nortel Networks, and Quarry Technologies. Specifically, we intend to use Lucent's Network Element for Programmable Packet Injection (NEPPI). NEPPI provides an ideal foundation upon which to implement the control policies we propose to develop. This project is a collaborative efforts between Boston University (Is: Ibrahim Matta, Azer Bestavros, and Mark Crovella) with expertize in characterization, measurements and control of Internet traffic, and University of Arizona (PI: Marwan Krunz) with expertize in traffic modeling, multimedia and wireless QoS.