Thèse et Dissertations

  1. Steffenel, L.A. "Contributions à la Gestion de l’Hétérogénéité dans les Environnements Distribués et Pervasifs", HDR thesis, Université de Reims Champagne-Ardenne, 2017. PDF Slides
     
  2. Barchet-Steffenel, L.A. "LaPIe : Communication Collectives Adaptées aux Grilles de Calcul". PhD Thesis, INPG, Grenoble, France, December 2005. PDF BibTeX
     
  3. Barchet-Steffenel, L.A. "Analyzing RBP, a Total Order Broadcast Protocol for Unreliable Channels". MSc. Dissertation, Doctoral School in Communication Systems, EPFL, Lausanne, Switzerland, July 2002. PDF BibTeX
     
  4. Barchet-Steffenel, L. A. "Avaliação dos Detectores de Defeitos e sua Influência nas Operações de Consenso" (Evaluation of Failure Detectors and their Influence on the Consensus Operations), Master Thesis, PPGC:UFRGS, Porto Alegre, Brazil, March 2001. PDF BibTeX
     
  5. Barchet-Steffenel, L. A. "CsocketS - Classes para Comunicação de Rede com Autenticação e Criptografia usando Java e CORBA" (CSocketS - Classes for Network Communication with Authentication and Criptography, using Java and CORBA). Bsc. Dissertation, UFSM, Santa Maria, Brazil, March 1999. PDF Source Code

BibTeX


PhD Thesis - LaPIe - Communications Collectives Adaptées aux Grilles de Calcul

 

Thèse labelisée "C3I" - Certificat de Compétences en Calcul Intensif par le GENCI (Grand Équipement National de Calcul Intensif) et la CPU (Conférences des Président d’Universités)



 

 
  • Directeurs de Thèse: Prof. Denis TRYSTRAM et Prof. Grégory MOUNIE
     
  • Date de Soutenance: le 2 Décembre 2005
     
  • Résumé : Avec la démocratisation des environnements du type grappe et grille de calcul, la performance des opérations de communication collective devient un aspect critique dans le coût total des applications parallèles. Fortement influencées par l'hétérogénéité des ressources, ces opérations dépendent à la fois des paramètres de communication des réseaux et des stratégies de communication employées.
    Cette thèse a pour objet d'étude l'optimisation des communications collectives selon l'approche préconisée par Karonis, où les différentes grappes de calcul sont organisées en plusieurs couches de communication de manière à minimiser le temps total de communication. Pour cela, nous proposons des modèles de communications qui permettent l'optimisation des communications collectives à travers l'ordonnancement hiérarchique des communications et la prédiction des performances.
    Dans un premier temps, nous démontrons que l'utilisation du modèle de coût pLogP permet la modélisation des performances de communications collectives dans des environnements homogènes. Ces modèles de performance ont été validés expérimentalement sur différentes plates-formes réseaux (Fast Ethernet, Giga Ethernet et Myrinet).
    Parallèlement, nous étudions la découverte automatique de la topologie du réseau. En effet, la décomposition de l'environnement de grille en îlots d'homogénéité permettrait la réduction de la complexité des optimisations, notamment dans ce qui concerne la modélisation des performances et l'obtention des caractéristiques du réseau.
    Notre principale contribution a été de proposer certaines heuristiques d'ordonnancement des communications qui tiennent compte de l'organisation hiérarchique de la grille. Ces heuristiques, adaptées aux différents patrons de communication collective, utilisent les techniques étudiées précédemment (dont la découverte des îlots d'homogénéité et la modélisation des performances) afin de réduire la complexité de l'optimisation des communications et de minimiser le temps total de communication.
     
  • Mots clés: communication collective, modélisation de performance, grilles de calcul, MPI

MSc in Communication Systems

Doctoral School in Communication Systems
Ecole Polytechnique Fédérale de Lausanne (EPFL)
Lausanne, Switzerland.

  • Advisor: Professor André Schiper
  • Dissertation: “Analysing RBP, a Total Order Broadcast Protocol for Unreliable Channels
  • Abstract: RBP is the acronym for Reliable Broadcast Protocol. It was published by Chang and Maxemchuck in 1984, and aims to provide total order and reliable broadcast in a distributed system subjected to process failures and fair lossy links. While the original definitions are not recent, new protocols and implementations still make use of its concepts (for example, RMP, Pinwheel and TRMP). We started our analysis by comparing RBP with another Total Order solution, the Primary Backup+VSC replication model. As View Synchronous Communication (VSC) is widely studied, it presents many techniques that can be used in order to update RBP. One of these techniques, called "two views" model is specially interesting because it allow us to take advantage from failure detectors with aggressive timeouts while minimizing the drawbacks from incorrect suspicions. Based on the analysis we ve done, we propose an improved version from the RBP protocol that uses the techniques studied in this work.

MSc in Computer Science

Graduate School in Computer Science
Universidade Federal do Rio Grande do Sul (UFRGS)
Porto Alegre, Brazil.

  • Advisor: Professor Ingrid Jansch-Pôrto
  • Dissertation: “Evaluation of Failure Detectors and their Influence on the Consensus Operations
  • Abstract: This work presents our observations and analysis on the influence of the failure detectors on the consensus algorithm. Failure detectors are essential to the consensus over an asynchronous distributed system, as they represent one of the few techniques that are able to circumvent the limitation imposed by the FLP Impossibility (the impossibility to distinguish a crashed process from a slow one, in asynchronous systems). While failure detectors are well defined through two properties, completeness and accuracy, there's no rule about their implementation. Indeed, we only found one work that tried to compare different failure detectors. Using simulation techniques, that work evaluated the impact of the failure detectors on the consensus termination time, although that research was very limited in the number of detectors analyzed and in the evaluation goals. The present work extended that experience, including more detectors in the analysis and evaluating them in a practical environment. Also, the observations were not restricted to those from the original paper, and the detectors were analyzed with more metrics, failure patterns and operational parameters. The evaluation allowed us to identify the behavior from the detectors in face of the most significant failure patterns, their influence on the consensus operation and their interaction with the execution environment. This evaluation also enabled us to compare the detectors, identifying their limitations, their best employment situations and possible optimizations to future developments.

BSc in Computer Science

Universidade Federal de Santa Maria (UFSM)
Santa Maria, Brazil.

  • Advisor: Professor João Carlos Damasceno Lima
  • Dissertation: “CsocketS - Classes para Comunicação de Rede com Autenticação e Criptografia usando Java e CORBA
  • Abstract: Security is a basic requirement for the systems developed to the Internet. The set of classes for network communications with authentication and encryption _ CsocketS _ have as goal to offer an API that could be easily integrated to the programming code from existing applications, extending the functionalities of sockets through the use of encryption, digital signature and session authentication. The system was developed using Java and CORBA, guaranteeing easiness of distribution and the interface homogeneity. We also propose a proxy architecture, enabling the handling to different applications with the use of one same server and to work through firewalls.