ED Mathématiques et Informatique
Optimizing HPC Applications with Vectorization and Multi-Criteria Task Scheduling on Heterogeneous Systems
by Hayfa TAYEB (LaBRI - Laboratoire Bordelais de Recherche en Informatique)
The defense will take place at 16h00 - Salle Ada Lovelace (Etage 3, Salle A303) Centre de recherche Inria de l'université de Bordeaux 200 Avenue de la Vieille Tour 33405 Talence
in front of the jury composed of
- Abdou GUERMOUCHE - Maître de conférences - Université de Bordeaux - Directeur de these
- Julien LANGOU - Professor - University of Colorado Denver - Rapporteur
- Pierre FORTIN - Professeur des universités - Université de Lille - Rapporteur
- Florina M. CIORBA - Associate Professor - University of Basel - Examinateur
- Vicenç BELTRAN - Directeur de recherche - Barcelona Supercomputing Center - Examinateur
- Raymond NAMYST - Full professor - Université de Bordeaux - Examinateur
- Bérenger BRAMAS - Chargé de recherche - Inria - CoDirecteur de these
High-performance computing (HPC) applications, such as numerical simulations, have huge computing requirements. These applications rely on software that effectively leverages hardware capabilities. This thesis focuses on optimizing HPC applications by intervening at different levels of the software stack. The optimization targets various metrics: performance (GFlop/s), energy efficiency (GFlop/s/Watt), and execution time (s). The emergence of heterogeneous HPC architectures has increased computational power but created challenges in software utilization. Task-based runtime systems offer a solution, where scheduling is critical for efficient resource utilization. We propose a new multi-criteria scheduler and evaluate task-based applications in real-world scenarios with StarPU runtime system. Challenges in dynamic task scheduling are addressed, emphasizing their impact on accelerating HPC applications. While higher performance remains a priority, it comes at the cost of increased power consumption. Energy efficiency provides an interesting way to balance power consumption and computational performance. In this thesis, we propose to study the impact of GPU power capping in the context of HPC applications using heterogeneous computing systems. We investigate how setting different power caps for GPU devices can improve the energy efficiency of the running application. At a lower level of the software stack, optimizations intervene at the level of the kernel code. Data-level parallelism, achieved through vectorization, is key to maximizing computational speed. We introduce Autovesk, a source-to-source code generation tool for vectorizing static kernels. Autovesk overcomes the limitations of existing tools, efficiently handling non-contiguous data access patterns that prevent the automatic vectorization by the compiler.
ED Sciences Chimiques
Valorization of biobased tartaric acid and film-forming biopolymers for the biocontrol of phytopathogenic and ice-forming microorganisms
by Solène MEYNAUD (Laboratoire de Chimie des Polymères Organiques)
The defense will take place at 9h30 - Amphithéâtre du CRPP Centre de Recherche Paul Pascal 115 Avenue du Dr Albert Schweitzer, 33600 Pessac
in front of the jury composed of
- Véronique COMA - Maîtresse de conférences - Université de Bordeaux - Directeur de these
- Cindy MORRIS - Directrice de recherche - INRAE Unité de Recherche de Pathologie Végétale - CoDirecteur de these
- Fabrice COLOMA - Docteur - Entreprise Faure - Examinateur
- Fernando LEAL-CALDERON - Professeur des universités - Bordeaux INP - Examinateur
- Frederic DEBEAUFORT - Professeur des universités - Université de Bourgogne - Rapporteur
- Joël POTHIER - Senior Researcher - ZHAW Zurich University of Applied Sciences - Life Sciences and Facility Management - Rapporteur
European fruit production is very affected by diseases caused by bacteria, such as Pseudomonas syringae or Erwinia amylovora. In addition, some governments have put in place restrictions on the use of phytosanitary products. In France, it corresponds to the Ecophyto Plan. Thus, many alternatives have been found to replace the phytosanitary treatments that present the most risk. However, one of the major challenges for the use of these compounds is rainfastness, to guarantee efficiency in the field over time. Among the bio-based compounds that are candidates for the treatment of plant pathogenic bacteria, tartaric acid shows high availability and low cost, but poor rainfastness due to its high solubility in water. This thesis focuses on the use of tartaric acid in rain-resistant and bioactive coatings against P. syringae and E. amylovora. To this end, formulations were developed using a chitosan solution as a film-forming matrix, tartaric acid as a pH regulator, and stabilized hydrophobic particles (zein or tristearin) as adjuvants to enhance rain resistance. Particles were first characterized, and the formulations were evaluated in terms of coating ability, water resistance, and bioactivity. In vivo tests were then conducted on plants using different methodologies to identify the most effective system. Additionally, a preliminary study was carried out to assess the interactions between tartaric acid and the plant, as well as the resistance of the coatings to rain, in order to predict the persistence of the bioactive compounds. In vivo tests were then conducted on plants using different methodologies to identify the most effective system. Additionally, a preliminary study was carried out to assess the interactions between tartaric acid and the plant, as well as the resistance of the coatings to rain, in order to predict the persistence of the bioactive compounds. Finally, the mechanisms of action were explored based on concentration, exposure time, and pH, to determine the extent of the formulation effectiveness and optimize their performance.
ED Sciences Physiques et de l'Ingénieur
Physically based modeling of the colored appearance of complex media for cosmetics
by Quoc TRAN (Laboratoire Photonique, Numérique & Nanosciences)
The defense will take place at 14h00 - Amphithéâtre Institut d'Optique Graduate School Nouvelle Aquitaine. 1 Rue François Mitterrand, 33400 Talence
in front of the jury composed of
- Anne PILLONNET - Professeure des universités - Université Claude Bernard Lyon 1 - Rapporteur
- Lionel SIMONOT - Maître de conférences - Université de Poitiers - Rapporteur
- Laurent BRUNEL - Ingénieur de recherche - PhotonLyX Technology SL - Examinateur
- Olivier MONDAIN-MONVAL - Professeur - Université de Bordeaux - Examinateur
- Philippe LALANNE - Directeur de recherche - CNRS - Directeur de these
Foundation is a colored, semi-transparent makeup product used to even out the complexion, conceal imperfections, or subtly alter skin tone. At the microscopic level, foundation is a complex and heterogeneous medium, generally composed, in its liquid form, of two phases (oil and water), solid particles (fillers, pigments, etc.) randomly distributed, and surfactants to ensure the stability of the formula. Establishing a relationship between the microscopic properties of such complex media and their visual appearance is a major challenge for digital formulation in cosmetics. This would allow, for example, to correct a priori variations in foundation ingredients or to predict the required composition for a specific targeted appearance. This thesis work contributes to this long-term objective by proposing a modeling approach for the prediction and analysis of the optical properties of complex heterogeneous media based on their composition and structure. An experimental study is first conducted on samples based on a simplified foundation formula to characterize their microscopic and optical properties. The foundation model used is that of a pigmented emulsion, namely a solution containing both solid, sub-micrometer particles and spherical, homogeneous, and transparent droplets of several tens of micrometers in size. A modeling tool based on a Monte Carlo method is developed to simulate light transport in such media, considering the scattering and absorption phenomena produced by the particles as well as the refraction and reflection phenomena produced by the droplets. This allows us to show how variations in microscopic parameters, such as the nature and density of particles, as well as the size and density of droplets, impact the coloration of a macroscopic material in reflection. In a second step, we propose an improvement of the method, based on an implicit description of the light interaction events with the microstructure, to simulate more efficiently the light transport in such media. This thesis work offers an increased understanding of the microscopic origin of the visual appearance of foundations, highlighting the importance of considering the heterogeneity of the materials due to the dispersed phase.