Catálogo de publicaciones - libros

This paper focus on optimal Call Admission Control (CAC) policy for an IEEE 802.16 Wireless MAN. This policy has two objectives: (i) statistically guarantee the QoS of UGS, rtPS and nrtPS connections, (ii) maximize the average revenue of the wireless link. To find such optimal policy, we model our CAC agent as a Constrained Semi-Markov Decision Process (CSMDP). To the best of our knowledge, the only algorithm able to compute the optimal control policy of a CSMDP is based on the Linear Programming (L.P.) approach. Unfortunately, a realistic CAC problem with a large states space is intractable with the L.P. algorithm. Our work presents two contributions. First, the proposition of an optimal CAC for Triple-Play services support over a 802.16 WMAN. Second, the presentation of an alternative iterative algorithm that can be used to overcome the difficulties faced by the L.P. approach.

Attaining stability guarantees using distributed scheduling policies is an important design goal in multi-hop wireless networks. Several recent results have recently characterized tradeoffs tradeoffs between computation times and stability guarantees in multi-hop wireless networks. We summarize existing results that have substantially advanced the state of the art in this context, and discuss problems that remain open.

A control model over a finite horizon is considered, where the state process is not observable and has to be estimated with an observation process, where each state of the observation process represents a group of states of the unobservable process. We show how the model with partial information can be transformed in one with complete information with the help of the filter technique and conditional probabilities. As a main result we prove the connection between the original and the reduced model and we show an explicite representation of the conditional probability.

Admission control in queueing networks is often based on partial information on the network state. This paper studies how the lack of state information affects performance by considering a simple model for admission control. The model is analyzed by studying a related censored process that has a matrix-geometric steady-state distribution. Numerical results show how partial information may cause some performance characteristics in queueing networks to be nonmonotone with respect to service rates.

This paper addresses the problem of designing tractable dynamic admission control and/or routing policies in a Markovian model of parallel multi-server loss queues with reneging, which seek to optimize performance objectives of concern. Such problems are relevant in a variety of systems that provide distributed telecommunication or computing services. The paper shows the direct applicability of the author’s results in Niño-Mora (2002) [Dynamic allocation indices for restless projects and queueing admission control: a polyhedral approach. Math. Program. 361–413], where index policies based on restless bandits were developed in a broader setting. A well-grounded and tractable index policy for admission control and/or routing is thus proposed for the model of concern. Results of preliminary computational experiments are reported, showing that the proposed index policy is nearly optimal, and substantially outperforms conventional benchmark policies in the instances investigated.

A brief overview of some interesting dynamics commonly arising as scaling limits of stochastic approximation type algorithms is given, with sample applications to communications.

We discuss the design of an overload control protocol from an optimisation perspective. The design process starts with the formulation of a convex optimisation problem. We then construct a distributed algorithm for this problem by applying Lagrangian decomposition. The different components make use of models of their environment to ensure convergence of the algorithm. We show how the constructed algorithm is implemented by GOCAP, an overload control protocol currently undergoing standardisation.

Many network control problems can be formulated and studied using the machinery of optimisation theory and Lagrange duality. The goal of the control process is to find the saddle point of the Lagrangian. We present a stability result for a class of dynamic processes for this problem. Our formulation automatically derives a Lyapunov function from the form of the dynamic equations. We show how several stability results from the literature of distributed flow control in networks fit into this formalism.

In this paper we propose NCRS, a Network Computational Resource Sharing grid based on network RAM. In NCRS, the computing node regards the free memory of other nodes in networks as a complement of local memory and uses it to store the large amount of intermediate data during computation. When the remote data is required, rather than blocks and gets it from remote nodes (or local disks), the computing node sends the related instructions to the remote nodes where the data locates. We refer to the memory on the remote nodes in networks as Network Intelligent Memory (NIM). NIM carries out the received instructions, and by this means the computing node reduces the page-swaps with local disks and the instructions locally carried out.

Optical Burst Switching proposes a future-proof alternative to the current electronic switching in the backbone. The involved optical buffers are implemented with a set of Fiber Delay Lines, and suffer serious performance loss, when compared to RAM. Various existing models trace this loss, but either lack generality, accuracy, or effectiveness.

The optical buffer model we constructed is valid for general line lengths and burst sizes. An effective approach allowed to strongly reduce the solution’s complexity, while remaining exact. This document presents the key formulas and performance graphs. The obtained model serves as a basic optimization tool, yielding results fast.