A family of QoS-aware traffic control protocols.

摘要

As the Internet has evolved into a global commercial infrastructure, there has been a growing demand for Internet to support various real-time applications, such as voice over IP, IPTV and multimedia streaming. These applications call for various real-time services including Hard Real-time (HRT), Real-time Delay Adaptive (RDA) and Real-time Rate Adaptive (RRA) services. However, the existing transport layer protocols, including TCP and User Datagram Protocol (UDP), cannot provide such services. In today's Internet, the QoS features can only be enabled through the help of the core network nodes using technologies such as Resource ReSerVation Protocol (RSVP), DiffServ or Multiprotocol Label Switching (MPLS). These approaches, however, require significant involvement of the core network nodes and therefore have difficulty to scale to the global Internet. Hence it is necessary to design better distributed end-to-end congestion control mechanism to enable QoS feature for real-time applications.;In this thesis, we design a family of end-to-end traffic control solutions for both the reliable and the unreliable services.;Firstly, we propose the end-to-end traffic control design methodology, known as TCP-Elastic Real-time SErvice (TERSE). TERSE provides a unified end-to-end traffic control protocol that enables the Non-Real-time Elastic (NRE) service (i.e., the same as the one under the TCP control), Real-time Delay Adaptive (RDA) service, and Real-time Rate Adaptive (RRA) service. A specific service is enabled by properly setting a single parameter in the protocol only. TERSE is underpinned by a sound design methodology. While having its root in the well-known utility-based optimization approach, this methodology successfully addresses the limitations of the utility-based optimization approach. It leads to the successful design of the unified traffic control protocol, which enjoys some provable properties, including fairness, convergence, and stability. Other than the traditional utility-based approach which expresses the QoS features implicitly in the utility function, TERSE methodology defines the utility functions of all the QoS services the same as TOE utility function and expresses the QoS features explicitly as boundary conditions of the optimization problem.;TERSE methodology requires the utility function of the prevailing TCP of Internet to design the end-to-end QoS aware protocols. So we derive a global utility function and the corresponding optimal control law, known as TCP control law, which maximizes the global utility. The TCP control law captures the essential behaviors of TCP, including slow start, congestion avoidance, and the binary nature of congestion feedback in TCP. We find that the utility function of TCP is linear in the slow start phase and is proportional to the additive increase rate and approaches the well-known logarithm function as the data rate becomes large in the congestion avoidance phase. We also show the fact that understanding the slow start phase with a fixed threshold is critical to the design of new transport layer control protocols to enable quality of service features.;For reliable services, based on TERSE and TCP utility function, we derive a family of optimal, distributed, QoS-aware, end-to-end congestion control laws. It enables a set of class of services (CoSs) including Assured Forwarding Service (AF), Minimum Rate Guaranteed Service (MRG), Upper Bounded Rate Service (UBR), and Minimum Rate Guaranteed and Upper Bounded Rate Service (MRGUBR). Also we use the fluid model to study the MRG control law under what conditions that it can achieve the target rate in competing with the TCP flows. These control laws are implemented as the unified window-based congestion control protocols, similar to the window-based TCP congestion control protocol. The performance of these protocols is tested based on NS-2 simulation. The results indicate that the protocols are indeed TCP friendly and can provide end-to-end service assurance as long as the percentage of network bandwidth consumed by the flows using these protocols is moderately small. Both analytical and simulation studies show that they can provide required soft minimum-rate guarantee for both RRA and RDA traffic flows. The MRG protocol is also implemented in LINUX-based systems. The cross-pacific testing of this protocol with soft minimum rate guarantee shows that it can achieve more than 1.2Mbps throughput performance, 150% higher than TCP-reno and 50% higher than TCP cubic, which is considered to be the fastest variation of TCP.;For unreliable services, we first redefine the congestion indicator by the help of the Single Trip Time (STT). Then, we design an end-to-end Congestion Control Identifier (CCID) for Datagram Congestion Control Protocol (DCCP) to support RRA and RDA applications. The theoretical upper bound of guaranteed minimum rate is derived and verified through simulations. We also compare the theoretical drop ratio of proposed DCCP-QoS mechanism with that of the DCCP-TCP-like mechanism and verifies the result by NS-2 simulation. The experimental results show that the proposed mechanism can provide minimum rate guarantee for real-time applications and maintain a lower packet drop ratio.