Annie Gravey

Optimization of hierarchical routing protocols

Abstract This paper concentrates on hierarchical routing protocols and their optimization. In a hierarchical routing protocol the network is divided into independent routing areas so that a node has complete routing information on its partition, but only reduced routing information of the “outside world” (i.e. the other partitions). Such routing protocols have been introduced in the early 1980s in order to decrease the overhead of dynamic routing protocols in large networks. However, such an approach may degrade the network performance. The aim of this work is to understand and quantify some aspects of the influence of this clustering process on network performance, and to determine how a given network (of a given topology) should be clustered under some optimization criteria. This problem has been already considered in the past, using a deterministic worst-case approach, for different optimization criteria. We generalize and refine the known results using a random geometric approach. This framework allows us to calculate explicitly the averages of various relevant variables (such as the size of the routing table and connection set-up latency) as a function of the topological parameters (such as the number of hierarchical levels). Optimum structures are then easily deduced.

A survey of straightforward statistical multiplexing models for ATM networks

Connection Acceptance Control (CAC) is probably the most important function in the preventive congestion control strategy to be implemented in ATM networks. The CAC function must decide, on the basis of the traffic descriptors provided by an incoming connection, whether the new call can be accepted or not. A key factor in the access decision is the behaviour of the superposition of ATM VCs in a switch output buffer, and in particular the cell loss probability resulting from accepting the connection. Many models have been proposed in the literature in order to investigate this issue. In this paper, we present a survey of the queuing models used to estimate the cell loss probability in ATM networks with straightforward statistical multiplexing. The CAC policy under consideration is based on the so-called Worst Case Traffic allocation corresponding to the traffic descriptor parameters submitted by a VC: itspeak cell rate, cell delay variation tolerance, sustainable cell rate, andintrinsic burst tolerance parameters, when applicable. We plan to investigate the practical application of these models to the problem of CAC; in particular, the important issues of accuracy of the available models, their range of validity and their complexity deserve more attention than they have received up to now. The present survey stands as a first step in that direction.

Performance evaluation of an optical transparent packet switch

This paper presents first results concerning the performance evaluation of an optical transparent packet switch that solves contention using optical delay lines. In the switch under study, delay lines are grouped in several sets, as the first stage of the switch. Each input port has access to a few of these delay lines, and each set of delay lines has access to each output port. Non-FIFO output buffers are thus emulated using scheduling on a small number of delay lines with non-consecutive delays. Under simplifying assumptions, analytical models are derived, and validated by simulation. These models provide efficient bounds for estimating packet loss probability, under the assumption of regular, balanced input traffic. It is shown that the proposed switch architecture achieves a good performance in terms of packet loss, with a number of delay lines significantly smaller than the one currently used in other architectures.

Resource allocation for worst case traffic in ATM networks

Abstract This paper addresses the problem of resource allocation for ATM connections carrying either CBR or VBR applications, for the first generation of ATM equipments. Evaluating the resources needed for the transfer phase of an ATM connection is a first step in the design of a CAC taking into account both call and transfer level QoS. The required amount of resources is estimated on the basis of traffic characteristics negotiated at call set-up, under the assumption that the offered traffic is a purely periodic pattern of active and inactive periods; this assumption is currently considered to yield upper bounds for the amount of resources which must be allocated to a connection. The traffic characteristics, and the corresponding periodic patterns of traffic are obtained using the ITU-T specification of ATM traffic control. We identify three frameworks (allocation on the basis of PCR, burst loss and burst delay framework in case of statistical multiplexing) and principles on which resource allocation can be based in each of these frameworks are highlighted.

A congestion control mechanism for connectionless services offered by ATM Networks

We propose a traffic management mechanism for connectionless networks on top of ATM infrastructures. The mechanism combines flow control at the packet layer (connectionless layer) and dynamic bandwidth allocation of the ATM connections interconnecting the connectionless servers of the connectionless network. Optimal mechanisms are obtained through Markov decision processes for a model of two tandem queues. The obtained bandwidth gain motivates the analysis of such mechanisms in a more realistic model. The simulation of a more detailed model of a connectionless network allows us to conclude on the favorable impact of dynamic resource allocation on the bandwidth gain and on the reduction of the sensitivity of the performances of the network with respect to the characteristics of the traffic. The traffic management mechanism implemented in the simulator are motivated by the optimal mechanism obtained using the analytical model.

Models for setting ATM parameter values

In ATM networks, a user should negotiate at connection set-up a traffic contract which includes traffic characteristics and requested QoS. The traffic characteristics currently considered are the Peak Cell Rate, the Sustainable Cell Rate, the Intrinsic Burst Tolerance and the Cell Delay Variation (CDV) tolerance(s). The values taken by these traffic parameters characterize the so-called “Worst Case Traffic” that is used by CAC procedures for accepting a new connection and allocating resources to it. Conformance to the negotiated traffic characteristics is defined, at the ingress User to Network Interface (UNI) and at subsequent Inter Carrier Interfaces (ICIs), by algorithmic rules based on the Generic Cell Rate Algorithm (GCRA) formalism. Conformance rules are implemented by policing mechanisms that control the traffic submitted by the user and discard excess traffic. It is therefore essential to set traffic characteristic values that are relevant to the considered cell stream, and that ensure that the amount of non-conforming traffic is small. Using a queueing model representation for the GCRA formalism, several methods are available for choosing the traffic characteristics. This paper presents approximate methods and discusses their applicability. We then discuss the problem of obtaining traffic characteristic values for a connection that has crossed a series of switching nodes. This problem is particularly relevant for the traffic contract components corresponding to ICIs that are distant from the original source.

Performance Evaluation of the KEOPS Wavelength Routing Optical Packet Switch

This paper presents results concerning the performance evaluation of the KEOPS wavelength routing optical transparent packet switch. This switch solves contention using optical delay lines; these delay lines are grouped in several sets, in the first stage of the switch. Each input port has access to a few of these delay lines, and each set of delay lines has access to each output port. Non-FIFO output buffers are thus emulated using scheduling on a small number of delay lines with non-consecutive delays. Under simplifying assumptions, analytical models are derived, and checked by simulation. These models provide efficient bounds for estimating packet loss probability, under the assumption of regular, balanced input traffic. It is shown that the proposed switch architecture achieves a good performance in terms of packet loss, with a number of delay lines significantly smaller than the ones currently used in other architectures.

A model for evaluating the impact of aggregated routing information on network performance

This paper presents a mathematical model for evaluating the impact of information aggregation on the performance of PNNI-driven ATM networks. The routing aggregation scheme affects both the performance and scalability of these networks. However, to date, little is still known on how the choice of PNNI configuration parameters influences the network behavior. A generic model for the aggregation process is proposed in order to study its impact on the network utilization and routing overhead. Random geometric considerations are used to obtain closed-form approximations of various performance measures.

Dynamical Resource Reservation Scheme in an ATM Network Using Neural Network-Based Traffic Prediction

Using real traffic data, we show that neural network-based prediction techniques can be used to predict the queuing behaviour of highly bursty traffics typical of LAN interconnection in a way accurate enough so as to allow dynamical renegotiation of a DBR traffic contract at the edge of an ATM network.

Traffic management for connectionless services over ATM

This paper is concerned with the transport of connectionless traffic over ATM networks. Several architectures have been defined and standardized for this purpose, like LANE, CBDS, Classical IP or MPOA. However traffic management for resources optimization of these architectures remains mostly an open issue. One of the key problem there is to control – and dimension – the capacity of the required ATM virtual circuits. In [19,21], dynamic capacity allocation schemes have been defined for this purpose. They have shown great promise through simulation and analytical studies. In this paper, we analyze the behavior of a multiplexer with such dynamic capacity allocation schemes when fed by several bursty flows. EATA properties [5,18] are used to compute the loss probability seen by each individual multiplexed stream, extending the analytical analysis of [9]. The study gives insight into the usefulness of these schemes and highlights several interesting properties: namely, the reduction of inter‐dependency between multiplexed flows and the reduction of the average queue occupancy. Such properties allow a better multiplexing gain. The particular case of dynamic allocation schemes for TCP/IP is then discussed.