Jordi Ros

A theory of convergence order of maxmin rate allocation and an optimal protocol

The problem of allocating maxmin rates with minimum rate constraints for connection-oriented networks is considered. This paper proves that the convergence of maxmin rate allocation satisfies a partial ordering in the bottleneck links. This partial ordering leads to a tighter lower bound for the convergence time for any maxmin protocol. An optimally fast maxmin rate allocation protocol called the distributed constraint precedence graph (CPG) protocol is designed based on this ordering theory. The new protocol employs bi-directional minimization and does not induce transient oscillations. The distributed CPG protocol is compared against ERICA, showing far superior performance.

Minimum rate guarantee without per-flow information

This paper introduces a scalable maxmin flow control protocol which guarantees the minimum rate for each connection-oriented flow without requiring per-flow information. The protocol is called MR-ASAP (minimum rate guaranteeing adaptive source-link accounting protocol). MR-ASAP is an extension of ASAP, the first exact maxmin flow control protocol for best-effort connection-oriented traffic in integrated service networks, without requiring per-flow accounting at the intermediate network node. In the classical maxmin computation, only the maximum rate constraints are considered; in this paper the minimum rate requirements are treated similarly as the maximum rate constraints. Existing protocols that achieve exact maxmin optimality with minimum rate guarantee require per-flow information and complex computation such as sorting of the minimum rates at the switch. By generalizing the concept of constraint, the complex sorting and per-flow accounting required in the existing protocols are avoided. Simulation demonstrates fast convergence to optimality.

A general theory of constrained max-min rate allocation for multicast networks

This paper presents a general theory of network max-min rate assignment as a lexicographic optimization. The model includes multicast and lower bound constraints. The model for multicast allows the sender to send at the maximum rate allowed by the network and the receivers. Equivalent optimality conditions, especially those which can be coded into practical algorithms, are derived. A reference parallel algorithm is also derived. The theoretical results clearly show the important role of the advertised rates in automatable optimality conditions. The theory also shows that, once the single-link problem is solved, the multi-link (network) problem can be simply solved by recursively applying the algorithm for single-link problem.

Thresholds: Revisit the Strings Versus Clouds Debate for the Internet Architecture, Part II: QoS, Control, Management, and TCP

The clouds (IP) versus strings (connection-oriented) debate over the Internet architecture is reexamined. Controllability and observability are shown to be the key to the performance (QoS) of the networks. The clouds architecture treats the network as a black box, making it uncontrollable and unobservable; in contrast, the strings architecture was designed to be a controllable and observable structure. In network management, the need for centralized management and control to obtain efficiency and optimal performance argues for strings architecture. Finally, TCP is shown to be unscalable in performance because of its poor observability and controllability.

Thresholds: Revisit the Strings versus Clouds Debate for the Internet Architecture. Part I: Control, Scalability, and QoS

Debates over the Internet architecture again raise the issue of connectionless (clouds) architecture versus the connection-oriented (strings) architecture [1–5]. Quality-of-Service (QoS) technology is aimed at mediating between these two approaches, but there is little deployment of QoS systems. Why is QoS so hard to achieve? The answer is buried in the Clouds versus Strings debate. The Internet architecture is a multi-dimensional complex problem and it can be addressed from many different angles at many different levels of detail and perspective. This article focuses on the fundamental tradeoffs and optimizations in this debate; detailed investigations can be found in the DARPA-funded NewArch Project [3–5] and IETF. The end-to-end arguments and the datagram paradigm have served as the architectural model for the Internet since early 1980’s. The dominance of this design philosophy was first challenged by the ATM technology in mid 1990’s, and over the last few years, is again challenged by the resurgence of strings technologies such as MPLS [3]. In addition, many service and infrastructure providers have pushed many functions into the network. The deviations from the original end-toend design have become so large that even the original authors of this principle have to rethink its meanings and forms [3]. The clouds design philosophy [1] argues that an open, application independent (end-to-end), and distributed architecture is best suited for the performance, scalability, provisioning and ease-of-use of the Internet. For distributed architecture, reliability and scalability will be achieved by (a) separation of forwarding path and control plane; (b) pushing services out to the edge; and (c) globally meaning addresses in each packet to empower each node to make decisions. In particular,

Bottleneck branch marking for noise consolidation in multicast networks

The noisy feedback consolidation problem in point-to-multipoint ATM multicast networks is studied. A new algorithm, which keeps track of the M smallest available rates (AR) from the branches at each branching point, is proposed. This algorithm has zero response delay, noise stability (defined in the present paper), and small probability of noise. The probability model assumes no knowledge of the distribution of the available rate from the branches. Both analytical and simulation results demonstrate the superiority of the new algorithm.

A theory of temporal-spatial flow control: the case of single bottleneck link

This paper presents a theory of spatial-temporal flow control, the case of a single bottleneck link, for a network with dynamic available bandwidth. For this class of flow control, when the sources receive the feedback from the network, they schedule their rates considering both the amount of available resources as well as a timely allocation of these resources. With the theory developed in this paper, a protocol which ensures maximal utilization of the bandwidth, minimal queue size in the link and fairness among the sources is proposed. With this protocol, the sources estimate the amount of congestion they are producing at the bottleneck and the queue size at the bottleneck can be remotely controlled.

An optimal zero-queue rate control protocol for generalized MPLS networks

The problem of rate allocation and congestion control for generalized label switched networks such as those defined in the generalized MPLS model (for example, pure optical networks) is considered. First, rates have to be allocated to each flow in the discrete domain of labels. Second, due to an inconsistent forwarding problem, that we call the label reallocation routing problem, rates (labels) have to be dynamically assigned in such a way that network feasibility is ensured at all times. In other words, the flow control protocol has to work in an equivalent scenario where switches do not have queues. The paper proves that such zero-queue protocols exist, and we propose the first zero-queue protocol for the dynamic rate allocation problem of generalized label switched networks. The approach can be generalized to support a broader family of flow control protocols. This implies that, with the methodology proposed, traditional flow control protocols can be extended to support the zero-queue property.

A distributed multicast multi-rate maxmin flow control protocol

This paper presents the first distributed multi-rate maxmin rate allocation protocol for multicast traffic. The protocol supports minimum rate guarantee for each receiver as a QoS parameter. The protocol is practical and it maintains a good performance-scalability trade-off. While the network is required to participate, the complexity added by the protocol is minimal.