WP 1

WP1: Global performance metrics and macro-models for digital communication systems (all partners).

Digital communication systems and networks target the transmission of information with the highest possible reliability. Along the years, digital communication systems have become more and more evolved. Elements of complexity are a. o. the need for such systems to simultaneously serve multiple users having different QoS requirements (e.g. information rate, priority and delay expectations, packet error/loss, ...) and/or the complexity of the channels (wired, wireless) over which transmission has to take place. Digital communication systems have subsystems where parameters (e.g. channel, modulation, coding, scheduler, ) have to be designed and/or resources (e.g. power, frequencies, time slots, antennas, buffer space) have to be shared. A fully centralized design of the subsystems would lead to a much too complex problem. Therefore, the design of parameters and/or the resource sharing has to be implemented in a distributed manner, at the level of subsystems having a limited view on the global system. This has among others led to performance metrics that are subsystem-specific.

The aim of WP1 is twofold. First, performance metrics should be identified, which take into account as much as possible the different concerns expressed before (information rate, error rate, latency, energy, etc) and which are more global and representative of several of the subsystems considered. Second, attention will be paid to tools and models making it possible to convey an abstracted description of some subsystems, for the proper inclusion of this description in the investigation and optimization conducted for other subsystems.

SWP1.1: Cross-layer performance metrics for centralized and for distributed strategies.

SWP1.1 will be devoted to the identification of performance metrics accounting as much as possible for the different concerns that are information rate, (bit, symbol or packet) error rate, packet loss, delay, fairness, energy efficiency and complexity.

A performance metric associating several concerns, that will receive due attention in this project, is the weighted sum goodput. Goodput (i.e. the number of information bits delivered without error to the user by unit of time) has been proposed as a performance metric to account for information rate and error rates in packet oriented systems ([Qiao2002]). While it is often used as an a posteriori assessment metric, it is only recently that attempts were made to analytically optimize some link or system parameters for the goodput ([Devillers2008], [Cui2009], [Wang2011]). Weighting is a usual way to account for different priorities in the different links considered jointly. Another way to account for fairness is the max-min goodput criterion that will also receive due attention.

Because goodput as such may however not easily lead to closed form expressions for some kinds of problems, some macro-models may be used instead ([Stupia2009]). These macro-models will be investigated for the new goodput oriented problems considered in this project and will duly take HARQ into account ([Rui2008]).

In order to obtain an energy aware performance metric, it has been proposed to consider goodput per joule ([Meshkati2005]). Moreover, as mentioned before, a new challenge is to account not only for transmission power but also for computation power (named decoding power in ([Grover2011])). Therefore, effort will be devoted to obtain a performance metric combining goodput and total energy consumed, with the goal that it be usable by WPs 2 to 4.

SWP1.2: Stochastic macro-models for communication subsystems.

This SWP aims at developing macro-models for the subsystems investigated in WPs 2, 3 and 4. These macro-models should be simpler than the refined models originating from these WPs, capture the most important effects and be able to render the proper sensitivity to the most important parameters. This SWP will therefore focus on techniques enabling a sensitivity analysis ([Chemmangat2010], [Saltelli2008]) and a robust design ([Beyer2007]) of communication subsystems, in a multi-objective way. Multiple performance measures can be taken into account using Pareto optimization.

First, the relationship will be studied between selected input quantities and chosen output measures. The input quantities can be channel parameters (e.g., layout or material parameters for backplanes, scattering richness, multi-user separation for wireless systems, etc ), traffic traces (e.g. in video applications) and system inputs (e.g. allocation of bits, power and code rate to carriers, buffer space, etc ). The output measures are actually selected performance metrics (see SWP1.1) (e.g. power efficiency, bit error rate, complexity, packet loss rate, latency, etc ) for the subsystem under consideration. This input/output relationship will be studied using advanced sequential experimental design techniques ([Crombecq2011]), and multivariate macro-modeling techniques ([Ferranti2009], [Ferranti2010], [Chemmangat2010]), which make a thorough sensitivity analysis feasible, e.g. based on Monte Carlo ([El-Moselhy2011]), or generalized Polynomial Chaos (gPC) analysis ([Xiu2010]).

Experimenting with a wide range of input values offers insights into the response of the subsystem under extreme situations. For instance, if the performance of the subsystem changes significantly for small perturbations of some of the input values, then it may be operating in a parametric sensitive region. These critical regions and the related parameters must be identified in a reliable way by means of a suitable sensitivity analysis. This can help the designer to detect potential design problems, and to modify the subsystem in such a way that the effects of critical parameters are significantly reduced.

Alternatively, the sensitivity analysis also helps to identify the critical parameters, which means that care should be taken in WPs 2 to 4 to characterize these specific parameters accurately. Moreover, it is very likely that some parameters will be critical for some performance metrics, and not for others.

Eventually, this methodology will help to design and optimize algorithms or subsystems which are robust against parameter changes. One possibility is, based on the multivariate macro-models and the sensitivity information, to select one or more robust design points (i.e. points in the design space that are rather insensitive to changes in extraneous factors) using surrogate-based multi-objective robust optimization techniques ([Forrester2008], [Gorissen2008], [Gorissen2010], [Couckuyt2010]). Note that the proposed approach can also deal with multiple, possibly conflicting performance goals (e.g., low BER versus minimum power consumption).

Another possibility is to design the subsystem for a chosen performance metric which is properly averaged over the distribution of the critical parameter(s), or which takes into account a worst case situation.