Date: Friday, October the 9th at 10AM
Place: Salle des Conseils Monod, ENS de lyon, (2nd floor)

Title:Causal Graphs applied to Network Science

Abstract:
Since the adoption of the Internet telecommunication networks world-wide, in the 90’s, the landscape of telecommunication networks has evolved very fast. The number of actors participating in its performance and the number of services it provides is in a constant evolution. Due to the heterogeneity in the technologies on which telecommunication networks rely and the number of services developed on top of the Internet networks, modeling and understanding its performance has become a complex problem. In such a scenario, changing one component or modifying one mechanism of a given telecommunication system represents an intervention that is both expensive and hard to revert. On the other hand, predicting the effect of a given change in a given mechanism influencing the telecommunication network performance is hard as many parameters need to be taken into account and controlled experiments are hard to implement in practice. To answer these challenges, our work proposes to adopt a causal approach to study the performance of applications running over the Internet network such as FTP or services such as the CDNs. Based on passive observations exclusively, a causal approach allows capturing the role of the different parameters influencing telecommunication networks performance. Such an approach also allows predicting how a service would react to a given change in the underlying network properties without the need of any additional experiment or active measurement. While able to overcome the challenges faced in the study of telecommunication network performance, the causal theory has some limitations in its implementations that prevents its use in the domain of telecommunication networks. The focus of our work is to show how to overcome the limitations in the existing implementations of the causal theory and how to implement such theory to study telecommunication networks. We first revise the existing assumptions made in the literature in order to adapt, in a second step, the existing methods for causal model and causal effect inference to the constraints of telecommunication networks. After validating our approach in an emulated environment, we present the studies of several real case scenarios where the causal approach shows its superiority over a non-causal (correlation based) approach and highlights the benefits and potential of such approach.

Date: Wednesday the 8th of July,
Place: ENS de Lyon, Amphi B

Title: Choosing amongst heterogeneous servers with applications to cloud computing

Summary : Many cloud systems such as the Amazon EC2 provide resources for clients to use. These resources are shared between many tasks. A natural model is to use a processor sharing. In this talk I will discuss new advances in understanding good policies to choose amongst heterogeneous servers in order to obtain low latencies for job execution.

We will first discuss the so-called Power-of-two rule in the homogeneous case of identical servers where routing to the least occupied server amongst two randomly chosen servers results in a very low server occupancy and a so-called propagation of chaos or asymptotic independence. The hetero- geneous case has not been treated in the literature. In the heterogeneous case we will see that the stability region for randomized routing is strictly included in the maximal stability region that can be achieved by state independent routing. Therefore the average sojourn of tasks can be longer in randomized routing in heterogeneous systems. When the system is stable we completely character- ize the steady-state behavior of the server occupancies and show that it exhibits super-exponential decay and asymptotic independence among servers. To overcome the reduction in the stability region we show that a combination of state independent routing (biased sampling) to a server class combined with JSQ within the class recovers the stability region as well as the benefits of small server occupancies. We conclude with the situation when the server speeds at different clusters are not known

The techniques are based on a mean field analysis. Joint work with Arpan Mukhopadhyay (Waterloo).

Biography: The speaker was educated at the Indian Institute of Technology, Bombay (B.Tech, 1977), Imperial College, London (MSc, DIC, 1978) and obtained his PhD under A. V. Balakrishnan at UCLA in 1983.

He is currently a University Research Chair Professor in the Dept. of ECE at the University of Waterloo, Ont., Canada where he has been since September 2004. Prior to this he was Professor of ECE at Purdue University, West Lafayette, USA. Since 2012 he is a D.J. Gandhi Distinguished Visiting Professor at the Indian Institute of Technology, Bombay, India. He is a Fellow of the IEEE and the Royal Statistical Society. He is a recipient of the INFOCOM 2006 Best Paper Award and was runner-up for the Best Paper Award at INFOCOM 1998.

His research interests are in stochastic modelling and analysis applied to complex networks and systems and in issues of network science.

Sorry, this entry is only available in French.

Résumé :
Entre causalités microscopiques et phénomènes émergents, l’analyse des systèmes complexes ne peut véritablement se passer d’une approche multi-échelles. Nous abordons en particulier deux problèmes : celui de la représentation (multi-échelles) de données structurées et celui de la prédiction (multi-échelles) des systèmes dynamiques. Premièrement, nous proposons d’exploiter des mesures classiques développées en théorie de l’information pour évaluer la pertinence des niveaux de représentation/prédiction en termes d’information et de complexité(gain d’entropie, divergence de Kullback-Leibler, Information Bottleneck). Le problème de la représentation/prédiction optimale est donc formalisé sous la forme d’un problème d’optimisation combinatoire à deux objectifs, visant à extraire et agréger l’information disponible au niveau microscopique pour apporter des éléments d’analyse macroscopique. Deuxièmement, afin d’engendrer des représentations exploitables en pratiques par les experts, nous garantissons leur cohérence avec les connaissances a priori du système en contraignant l’espace de recherche du problème d’optimisation à partir des propriétés structurelles du système. Si le problème d’optimisation correspondant est NP-complet en général, nous proposons néanmoins des algorithmes polynomiaux dans le cas de structures particulières (hiérarchies, ensembles d’intervalles, produit cartésien des deux structures). Cette approche est notamment appliquée à l’agrégation spatio-temporelle de données médiatiques pour l’analyse multi-échelles des relations internationales, et à la prédiction des dynamiques multi-échelles dans un modèle classique de diffusion d’opinion (Voter Model).