WorldCat Identities

Laboratoire de l'informatique du parallélisme (Lyon)

Overview
Works: 129 works in 139 publications in 2 languages and 139 library holdings
Roles: Other, Publishing director
Publication Timeline
.
Most widely held works by Laboratoire de l'informatique du parallélisme (Lyon)
Profiterole : Un protocole de partage équitable de la bande passante dans les réseaux ad hoc by Rémi Vannier( Book )

2 editions published in 2009 in French and held by 2 WorldCat member libraries worldwide

Les réseaux ad hoc ont ceci de différent des réseaux avec point d'accès qu'ils nécessitent encore moins d'infrastructure, puisque les entités qui les composent, plutôt que de communiquer exclusivement avec le point d'accès, sont des noeuds mobiles qui s'auto-organisent pour former un réseau. Du point de vue du concepteur de protocoles, cette topologie de réseau a deux implications : le fait que le réseau s'auto-organise impose de concevoir des mécanismes distribués, et la mobilité intrinsèque à ces réseaux implique que ces protocoles doivent être réactifs. C'est à ces réseaux que je me suis intéressé pendant ma thèse, et plus spécifiquement aux difficultés qu'y pose la norme IEEE 802.11, norme des réseaux sans fil plus connue sous le nom de WiFi. On observe que dans ces réseaux, 802.11 partage souvent la bande passante de manière très inégale, avec les conséquences que cela entraîne sur les services utilisant ces réseaux, mais aussi sur la stabilité des liens radio. La couche MAC de 802.11, initialement développée pour les réseaux avec points d'accès, est à l'origine de ces problèmes. Le résultat principal de mes recherches est une proposition de protocole, appelé Profiterole, dont l'objectif est de réduire ces inégalités de partage de bande passante. Partant du constat que les informations disponibles au niveau MAC, souvent trop locales, sont insuffisantes pour assurer un partage à la fois efficace et équitable, le protocole Profiterole réside dans la couche routage et tire parti d'informations sur la topologie du réseau. La nouveauté de l'approche réside essentiellement dans la volonté que j'ai eu de la rendre applicable en pratique. L'algorithme de calcul des débits optimaux est donc distribué, et a un faible surcoût réseau. Durant l'exposé, je présenterai cette solution, et je tâcherai de démontrer les faiblesses des approches de niveau MAC et les avantages d'une approche de couche 3. Enfin, mes travaux m'ont conduit à m'intéresser aux métriques de qualité de service dans les réseaux sans fil. Je montrerai que l'équité à court terme a un impact positif sur la satisfaction à long terme des utilisateurs des réseaux ad hoc
Contributions to the Analysis of scaling Laws and Quality of Service in Networks : Experimental and Theoretical Aspects by Patrick Loiseau( Book )

2 editions published in 2009 in English and held by 2 WorldCat member libraries worldwide

In today's context of rapid expansion of the Internet, deep understanding of the statistical properties of network traffic is essential for Internet Service Providers to offer users the best Quality of Service possible. A major breakthrough in that direction was the discovery in 1993 of the self-similarity of network traffic, followed up by the ON/OFF model proposed in 1997 that posits the heavy-tailness of flow-size distributions as a plausible origin of this property. While of great interest, such mathematical models always rely on necessary simplifying assumptions which can limit their practical applicability to real networks, in particular due to the complexity of the TCP protocol. In this thesis, we use a hybrid approach based on the combination of real traffic traces, controlled experiments and theoretical developments to address some open questions concerning network traffic properties and their impact on QoS. Our experiments are based on a large-scale controllable testbed and an efficient traffic capture system. Then, we first address issues related to aggregate network traffic: we extend previous long-range dependent models and we propose an estimator of the flow-size distribution's tail index under sampling. We also perform an empirical study of the impact of long-range dependence and heavy-tails on QoS. Finally, we turn to the packet-level traffic of one TCP source and show, using a large-deviation principle, that it can be finely characterized by a multifractal structure intimately related to the control mechanism AIMD, and naturally reproduced by Markov models
Scheduling and Dynamic Management of Applications over Grids by Ghislain Charrier( Book )

2 editions published in 2010 in English and held by 2 WorldCat member libraries worldwide

The work presented in this Thesis is about scheduling applications in computational Grids. We study how to better manage jobs in a grid middleware in order to improve the performance of the platform. Our solutions are designed to work at the middleware layer, thus allowing to keep the underlying architecture unmodified. First, we propose a reallocation mechanism to dynamically tackle errors that occur during the scheduling. Indeed, it is often necessary to provide a runtime estimation when submitting on a parallel computer so that it can compute a schedule. However, estimations are inherently inaccurate and scheduling decisions are based on incorrect data, and are therefore wrong. The reallocation mechanism we propose tackles this problem by moving waiting jobs between several parallel machines in order to reduce the scheduling errors due to inaccurate runtime estimates. Our second interest in the Thesis is the study of the scheduling of a climatology application on the Grid. To provide the best possible performances, we modeled the application as a Directed Acyclic Graph (DAG) and then proposed specific scheduling heuristics. To execute the application on the Grid, the middleware uses the knowledge of the application to find the best schedule
Arrondi correct de fonctions mathématiques : fonctions univariées et bivariées, certification et automatisation by Christoph Quirin Lauter( Book )

2 editions published in 2008 in French and held by 2 WorldCat member libraries worldwide

This thesis broadens the research space for the implementations of correctly rounded mathematical functions in floating-point arithmetic. Interest goes first to the functions xn and to the function xy. A novel approach for detecting the rounding boundary cases of xy is proposed. This innovation yields to a consequent speed-up of this function. Further, this thesis proposes automatic approaches for certifying a correctly rounded implementation of a function. A new algorithm for the safe computation of infinite norms is presented and made practically usable. Finally, algorithms are developed that allow for an automation of the implementation process of functions. A first realization in the Sollya tool permits generating and certifying the code for evaluating a mathematical function without human interaction. In the future, an integration of such techniques in a compiler can be considered
Réseau longue distance et application distribuée dans les grilles de calcul : étude et propositions pour une interaction efficace by Ludovic Hablot( Book )

2 editions published in 2009 in French and held by 2 WorldCat member libraries worldwide

Parallel applications allow to execute many tasks of the same application, on many processors in the same time. These applications were first executed on supercomputers, end then on clusters. Today, this last infrastructure is evolving toward grids. Grids, in this thesis, are defined as an interconnection of clusters by a long-distance network. Most of MPI applications are written with the MPI standard, which describe how processes communicate, by message passing. Initialy design for clusters, MPI is still used to program parallel applications in grids. While different problems have been resolved for communications in clusters, the long-distance network of the grid raises many problems. The main objective of this thesis is to study in detail the inteeractions between the parallel applications and the transport layer in the long-distance network of computing grids, and then to solve these problems. In this thesis, we study in detail the use of TCP by MPI applications. In order to reduce the impact of retransmissions and the impact of congestion window, we propose a communication layer between the application (MPI for example) and the transport protocol (TCP) that is autoùatically and transparently executed. The general principle is to introduce proxies at the interface between the local network and the long-distance network to differentiate communications. These proxies allows to put forward the split of TCP connections in order to avoid losses and retransmissions on the long-distance link. This mecanism also allows to keep the congestion window closer to available throughput on the long-distance network
A Study of Spilling and Coalescing in Register Allocation as Two Separate Phases by Florent Bouchez( Book )

2 editions published in 2009 in English and held by 2 WorldCat member libraries worldwide

The goal of register allocation is to assign the variables of a program to the registers or to spill them to memory whenever there are no register left. Minimizing the spilling is a difficult problem, tightly bounded with the colorability of the program. The SSA form simplifies the coloring by splitting variables : we discovered that it makes the interference graph chordal. Our first goal was to better understand from where the complexity of register allocation does come, and why SSA seems to simplify the problem, by revisiting to the original proof of Chaitin (1981). The difficulty comes from the presence of (critical) edges and the possibility to perform permutations of colors or not. We studied the spill problem under SSA and several versions of the coalescing problem. We used it to design new heuristics to better solve the coalescing problem, so that an aggressive splitting can be used beforehand. This led us to promote a better register allocation scheme : First, spilling to reduce the register pressure to the number of registers, possibly by splitting a lot; Then color the variables and perform coalescing to remove most of the added copies. This scheme is expected to perform well in an aggressive compiler. However, the high number of splits and the increased compilation time required to perform the coalescing is prohibitive for just-in-time (JIT) compilation. So, we devised a heuristic, called "permutation motion," that is intended to be used with SSA-based splitting in place of our more aggressive coalescing in a JIT context
Towards an SSA based compiler back-end : some interesting properties of SSA and its extensions by Benoit Boissinot( Book )

2 editions published in 2010 in English and held by 2 WorldCat member libraries worldwide

This thesis is articulated around three topics, all of them related to the SSA form. First, we explore the optimization of liveness algorithms when applied to program in SSA form. In particular we present an efficient algorithm, based on the structure of the program (the loops) and based on properties specific to SSA form. Next we present an intermediate representation which is a variant of SSA: the Static Single Information form (SSI). We clarify the different definitions for this form which appeared in the literature, pointing out the points where they diverge. Then we prove that the interference graph (the intersection graph of the live-ranges) of variables under SSI form is an interval graph. Finally, we propose a new approach to the problem of SSA destruction. Our method, while simpler than previous approach, gives us results comparable to more complex and not always proved approaches
Implementation of binary floating-point arithmetic on embedded integer processors : Polynomial evaluation-based algorithms and certified code generation by Guillaume Revy( Book )

2 editions published in 2009 in English and held by 2 WorldCat member libraries worldwide

Today some embedded systems still do not integrate their own floating-point unit, for area, cost, or energy consumption constraints. However, this kind of architectures is widely used in application domains highly demanding on floating-point calculations (multimedia, audio and video, or telecommunications). To compensate this lack of floating-point hardware, floating-point arithmetic has to be emulated efficiently through a software implementation. This thesis addresses the design and implementation of an efficient software support for IEEE 754 floating-point arithmetic on embedded integer processors. More specifically, it proposes new algorithms and tools for the efficient generation of fast and certified programs, allowing in particular to obtain C codes of very low latency for polynomial evaluation in fixed-point arithmetic. Compared to fully hand-written implementations, these tools allow to significantly reduce the development time of floating-point operators. The first part of the thesis deals with the design of optimized algorithms for some binary floating-point operators, and gives details on their software implementation for the binary32 floating-point format and for some embedded VLIW integer processors like those of the STMicroelectronics ST200 family. In particular, we propose here a uniform approach for correctly-rounded roots and their reciprocals, and an extension to division. Our approach, which relies on the evaluation of a single bivariate polynomial, allows higher ILP-exposure than previous methods and turns out to be particularly efficient in practice. This work allowed us to produce a fully revised version of the FLIP library, leading to significant gains compared to the previous version. The second part of the thesis presents a methodology for automatically and efficiently generating fast and certified C codes for the evaluation of bivariate polynomials in fixed-point arithmetic. In particular, it consists of some heuristics for computing highly parallel, low-latency evaluation schemes, as well as some techniques to check if those schemes remain efficient on a real target, and accurate enough to ensure correct rounding of the underlying operator implementations. This approach has been implemented in the software tool CGPE (Code Generation for Polynomial Evaluation). We have used our tool to quickly generate and certify significant parts of the codes of FLIP
Contribution to the Deployment of a Distributed and Hierarchical Middleware Applied to Cosmological Simulations by Benjamin Depardon( Book )

2 editions published in 2010 in English and held by 2 WorldCat member libraries worldwide

The results presented in this thesis deal with the execution of applications on heterogeneous and distributed environments: computing grids. We study, from end-to-end, the process allowing users to execute complex scientific applications. The contributions of this work are thus manifold. 1) Hierarchical middleware deployment: we first present an execution model for hierarchical middleware. Then, based on this model, we present several heuristics to automatically determine the shape of the hierarchy that would best fit the users' needs, depending on the platform it is executed on. We evaluate the quality of the approach on a real platform using the DIET middleware. 2) Graph clustering: we propose a distributed and self-stabilizing algorithm for clustering weighted graphs. Clustering is done based on a distance metric between nodes: within each created cluster the nodes are no farther than a distance k from an elected leader in the cluster. 3) Scheduling: we study the scheduling of independent tasks under resources usage limitations. We define linear programs to solve this problem in two cases: when tasks arrive all at the same time, and when release dates are considered. 4) Cosmological simulations: we have studied the behavior of applications required to run cosmological simulations workflows. Then, based on the DIET grid middleware, we implemented a complete infrastructure allowing non-expert users to easily submit cosmological simulations on a computing grid
Dealing with Elephants in the Internet : Towards a Flow-Based Internet Architecture by Dinil Mon Divakaran( Book )

2 editions published in 2010 in English and held by 2 WorldCat member libraries worldwide

Si l'on en croit les tendances actuelles de la croissance du trafic Internet, la demande en bande passante dépassera la capacité du système dans quelques années. Cela pose un certain nombre de problèmes, parmi lesquels la possibilité d'apporter de la qualité de service (QdS) au trafic utilisateur, la méthode traditionnelle de sur-approvisionnement n'étant plus envisageable. Dans ce contexte, l'approche flow-aware networking (FAN), proposée comme alternative aux architectures habituelles IntServ et DiffServ, emploie une différentiation de service implicite en utilisant des mécanismes tels que le contrôle d'admission, l'ordonnancement, le routage, etc., au niveau des flux, un flux étant un ensemble de paquets possédant le même groupe d'identifiants et localisés dans le temps.Par définition, l'approche flow-aware networking est bloquante en ce qui concerne les flux. La politique de contrôle d'admission rejette les nouveaux flux selon les conditions du réseau. On définit un réseau basé flux (flow-based networking) comme une version généralisée de l'approche FAN, non restreinte à un des mécanismes de QdS y étant défini, et surtout sans blocage de flux. Le but est d'être flexible en combinant différents mécanismes basés sur les flux, proposés indépendamment, pour l'ordonnancement, la gestion des tampons, le contrôle de la congestion, le routage, la comptabilité, l'échantillonage, etc. Dans un tel réseau, la performance peut être basée sur des métriques du niveau des flux. Cela correspond aux priorités des utilisateurs finaux pour qui le délai des petits flux, le débit des gros flux, le taux instantané de trafic en continu, etc., sont souvent plus importants que les métriques de QdS à l'échelle du paquet
Réalisabilité et paramétricité dans les systèmes de types purs by Marc Lasson( )

1 edition published in 2012 in French and held by 1 WorldCat member library worldwide

This thesis focuses on the adaptation of realizability and parametricity to dependent types in the framework of Pure Type Systems. We describe a systematic method to build a logic from a programming language, both described as pure type systems. This logic provides formulas to express properties of programs and offers a formal framework that allows us to develop a theory of realizability in which realizers of formulas are exactly programs of the starting programming language. In our framework, the standard representation theorems of Gödel's system T and Girard's system F may be seen as two instances of a more general theorem. Then, we explain how the so-called « logical relations » of parametricity theory may be expressed in terms of realizability, which shows that the generated logic provides an adequate framework for developping a general theory of parametricity. Finally, we show how this parametricity theory can be adapted to the underlying type system of the proof assistant Coq and we give an original example of application of parametricity theory to the formalization of mathematics
Fault-tolerant and energy-aware algorithms for workflows and real-time systems by Li Han( )

1 edition published in 2020 in English and held by 1 WorldCat member library worldwide

Cette thèse se concentre sur deux problèmes majeurs dans le contexte du calcul haute performance:la résilience et la consommation d'énergie.Le nombre d'unités de calcul dans les superordinateurs a considérablement augmenté ces dernièresannées, entraînant une augmentation de la fréquence des pannes. Le recours à des mécanismes detolérance aux pannes est maintenant critique pour les applications utilisant un grand nombre decomposants pendant une période de temps significative. Il est par ailleurs nécessaire de minimiserla consommation énergétique pour des raisons budgétaires et environnementales. Ceci est d'autantplus important que la tolérance aux pannes nécessite une redondance en temps ou en espace quiinduit un surcoût énergétique. Par ailleurs, certaines technologies qui réduisant la consommationd'énergie ont des effets négatifs sur les performances et la résilience.Nous concevons des algorithmes d'ordonnancement pour étudier les compromis entre performance,résilience et consommation d'énergie. Dans une première partie, nous nous concentrons surl'ordonnancement des graphes de tâches sujets à des pannes. La question est alors de décider quelletâche sauvegarder afin de minimiser le temps d'exécution. Nous concevons des solutions optimalespour des classes de graphes et fournissons des heuristiques pour le cas général. Nous considéronsdans une deuxième partie l'ordonnancement de tâches périodiques indépendantes sujettes à deserreurs silencieuses dans un contexte temps-réel. Nous étudions combien de réplicats sontnécessaires et l'interaction entre dates butoir, fiabilité, et minimisation d'énergie
Représentations des polynômes, algorithmes et bornes inférieures by Bruno Grenet( )

1 edition published in 2012 in French and held by 1 WorldCat member library worldwide

La complexité algorithmique est l'étude des ressources nécessaires -- le temps, la mémoire, ... -- pour résoudre un problème de manière algorithmique. Dans ce cadre, la théorie de la complexité algébrique est l'étude de la complexité algorithmique de problèmes de nature algébrique, concernant des polynômes.Dans cette thèse, nous étudions différents aspects de la complexité algébrique. D'une part, nous nous intéressons à l'expressivité des déterminants de matrices comme représentations des polynômes dans le modèle de complexité de Valiant. Nous montrons que les matrices symétriques ont la même expressivité que les matrices quelconques dès que la caractéristique du corps est différente de deux, mais que ce n'est plus le cas en caractéristique deux. Nous construisons également la représentation la plus compacte connue du permanent par un déterminant. D'autre part, nous étudions la complexité algorithmique de problèmes algébriques. Nous montrons que la détection de racines dans un système de n polynômes homogènes à n variables est NP-difficile. En lien avec la question « VP = VNP ? », version algébrique de « P = NP ? », nous obtenons une borne inférieure pour le calcul du permanent d'une matrice par un circuit arithmétique, et nous exhibons des liens unissant ce problème et celui du test d'identité polynomiale. Enfin nous fournissons des algorithmes efficaces pour la factorisation des polynômes lacunaires à deux variables
Bases of relations in one or several variables : fast algorithms and applications by Vincent Neiger( )

1 edition published in 2016 in English and held by 1 WorldCat member library worldwide

In this thesis, we study algorithms for a problem of finding relations in one or several variables. It generalizes that of computing a solution to a system of linear modular equations over a polynomial ring, including in particular the computation of Hermite- Padéapproximants and bivariate interpolants. Rather than a single solution, we aim at computing generators of the solution set which have good properties. Precisely, the input of our problem consists of a finite-dimensional module given by the action of the variables on its elements, and of some elements of this module; the goal is to compute a Gröbner basis of the module of syzygies between these elements. In terms of linear algebra, the input describes a matrix with a type of Krylov structure, and the goal is to compute a compact representation of a basis of the nullspace of this matrix. We propose several algorithms in accordance with the structure of the multiplication matrices which specify the action of the variables. In the case of a Jordan matrix, we accelerate the computation of multivariate interpolants under degree constraints; our result for a Frobenius matrix leads to a faster algorithm for computing normal forms of univariate polynomial matrices. In the case of several dense matrices, we accelerate the change of monomial order for Gröbner bases of multivariate zero-dimensional ideals
Linear logic, type assignment systems and implicit computational complexity by Erika De Benedetti( )

1 edition published in 2015 in English and held by 1 WorldCat member library worldwide

La complexité implicite (ICC) vise à donner des caractérisations de classes de complexité dans des langages de programmation ou des logiques, sans faire référence à des bornes sur les ressources (temps, espace mémoire). Dans cette thèse, nous étudions l'approche de la logique linéaire à la complexité implicite. L'objectif est de donner des caractérisations de classes de complexité, à travers des variantes du lambda-calcul qui sont typables dans de tels systèmes. En particulier, nous considérons à la fois une perspective monovalente et une perspective polyvalente par rapport à l'ICC. Dans le premier cas, le but est de caractériser une hiérarchie de classes de complexité à travers un lambda-calcul élémentaire typé dans la logique linéaire élémentaire (ELL), où la complexité ne dépend que de l'interface d'un programme, c'est à dire son type. La deuxième approche rend compte à la fois des fonctions calculables en temps polynomial et de la normalisation forte, à travers des termes du lambda-calcul pur qui sont typés dans un système inspiré par la logique linéaire Soft (SLL); en particulier, par rapport à l'approche logique ordinaire, ici nous abandonnons la modalité “!” en faveur de l'emploi des types stratifiés, vus comme un raffinement des types intersection non associatifs, afin d'améliorer la typabilité et, en conséquence, l'expressivité. Enfin, nous explorons l'utilisation des types intersection, privés de certaines de leurs propriétés, vers une direction plus quantitative que l'approche qualitative habituelle, afin d'obtenir une borne sur le calcul de lambda-termes purs, en obtenant en plus une caractérisation de la normalisation forte
Évaluation de la bande passante disponible et traitement du partage dans les réseaux sans fil multisauts basés sur le standard IEEE 802.11 DCF by Van Nam Nguyen( )

1 edition published in 2012 in French and held by 1 WorldCat member library worldwide

The first part of the thesis is devoted to the evaluation of the available bandwidth. This is a parameter that is often used for admission control in these networks. In the literature, there are many methods for estimating the available bandwidth on a wireless link. However, these methods do not take into account the concept of retransmission which is defined in IEEE 802.11 and have clear impacts on the available bandwidth. We propose therefore a new passive method called RABE (Retransmission-based Available Bandwidth Estimation). Our method is based mainly on local information that can be detected by each node. We also evaluate the unconditional collision probability of packets from hidden nodes via an analytical model. This probability allows us to estimate the average number of retransmissions, which is then used to evaluate other parameters used by RABE. The simulations performed with the simulator ns2.33 show that our method produces more accurate results than other existing solutions. In the second part, we focus on the distribution of idle periods of a node that is often used to evaluate the performance of wireless networks based on IEEE 802.11. The problem is that there are different assumptions for this distribution in literature. We first characterize different types of idle periods in some fundamental scenarios such as the scenario as asymetric hidden stations and the scenario of flow in the middle. We then generalize these types of idle periods for relatively highly loaded networks. The obtained simulation results in ns2.33 are similar to our proposed model. This work also brings out that these distributions are multimodal, unlike what is often assumed in the literature
Numerical Quality and High Performance In Interval Linear Algebra on Multi-Core Processors by Philippe Theveny( )

1 edition published in 2014 in English and held by 1 WorldCat member library worldwide

This work aims at determining suitable scopes for several algorithms of interval matrices multiplication.First, we quantify the numerical quality. Former error analyses of interval matrix products establish bounds on the radius overestimation by neglecting the roundoff error. We discuss here several possible measures for interval approximations. We then bound the roundoff error and compare experimentally this bound with the global error distribution on several random data sets. This approach enlightens the relative importance of the roundoff and arithmetic errors depending on the value and homogeneity of relative accuracies of inputs, on the matrix dimension, and on the working precision. This also leads to a new algorithm that is cheaper yet as accurate as previous ones under well-identified conditions.Second, we exploit the parallelism of linear algebra. Previous implementations use calls to BLAS routines on numerical matrices. We show that this may lead to wrong interval results and also restrict the scalability of the performance when the core count increases. To overcome these problems, we implement a blocking version with OpenMP threads executing block kernels with vector instructions. The timings on a 4-octo-core machine show that this implementation is more scalable than the BLAS one and that the cost of numerical and interval matrix products are comparable
New Encryption Schemes and Pseudo-Random Functions with Advanced Properties from Standard Assumptions by Radu Titiu( )

1 edition published in 2020 in English and held by 1 WorldCat member library worldwide

In this thesis, we study the security of advanced cryptographic primitives against adversaries that behave closer to real-life scenarios. Namely, they can adaptively update their strategy during the attack, based on previously obtained information, possible from external sources like corrupted users. We construct Distributed Pseudorandom Functions that still output random-looking values, even when the adversary can adaptively corrupt some servers. Such a system assumes that the secret key is shared among multiple servers that have to combine their partial evaluations in order to obtain a pseudorandom value. We also prove security against adaptive corruptions, in the stronger simulation-based security model, for Inner Product Functional Encryption. Such a public-key scheme encrypts vectors x and can issue multiple secret keys associated to key vectors y. The decryptor learns the partial information <x,y> but nothing else. This primitive can compute statistics (e.g., weighted sums or means) on a database, while keeping each individual input private. We also construct a labeled variant, wherein each database entry is encrypted by a different client, called Multi-Client Functional Encryption.We finally provide a new construction of Non-Interactive Zero-Knowledge proof, which convinces a verifier of the validity of some NP statement without leaking anything else. In addition, an adversary obtaining many simulated proofs for possibly false statements cannot produce a valid proof of its own for a false statement. This primitive is used as a building-block for public-key encryption schemes with advanced security properties
Data-driven Management Solution for Microservice-based Deep Learning Applications by Zeina Houmani( )

1 edition published in 2021 in English and held by 1 WorldCat member library worldwide

We live in a new era of Big Data, the era of insights. While our capacity to collect real-time data has grown significantly over the past decade, our ability to analyze that data to turn it into knowledge has not kept pace. With new generations of devices and network technologies, the focus of Big Data is shifting toward the design of tools and applications able to extract information from collected data. The majority of emerging applications present expectations of near-real-time processing to maintain the validity of their results. However, guaranteeing their performance requirements is hampered by the traditional Cloud system designs and management strategies. Current systems for Big Data applications rely on heterogeneous resources distributed across the constrained Edge and the powerful Cloud. In addition, the applications are now created as a set of self-contained microservices, developed by independent teams following the DevOps practices. This evolution of systems designs has introduced extreme heterogeneity and uncertainty into emerging applications, highlighting the limitations of traditional management strategies.In this thesis, we focus on designing a system for Big Data applications that rethinks existing management strategies with a particular emphasis on the heterogeneity of incoming data, applications, and resources. We first study the decoupling of data producers and consumers in emerging microservice-based applications as the entry point to effectively leverage available services, even newly published ones. Accordingly, we propose a data-driven service discovery framework based on data-centric service descriptions and rely on a Peer-to-Peer data-driven architecture. In addition, we present an adaptation scheme that scales deployed microservices to tackle the impact of fluctuating load on real-time performance. Second, we investigate the trade-off between the quality and urgency of the results in Big Data applications as a promising strategy to overcome the limited and heterogeneous capacity of system resources. In particular, we present a data-driven workflow scheduling approach to distribute microservices across the edge of the network, the core, and along the data path. Additionally, we propose a data adaptation strategy that reduces the quality of incoming data when potential quality-latency trade-off optimizations are available. We then apply the proposed approaches in the context of Deep Learning applications
 
moreShow More Titles
fewerShow Fewer Titles
Audience Level
0
Audience Level
1
  General Special  
Audience level: 0.00 (from 0.00 for Profiterol ... to 0.00 for Profiterol ...)

Alternative Names
Computer Science Laboratory

Ecole normale supérieure

Ecole Normale Supérieure de Lyon Laboratoire d'Informatique du Parallélisme

Ecole normale supérieure Lyon Laboratoire de l'informatique du parallélisme

Laboratoire de l'Informatique du Parallélisme facility in Lyon, France

Laboratoire d'informatique du parallélisme

LIP

UMR 5668

Languages
English (20)

French (10)