ED Mathématiques et Informatique
Dynamic scheduling for deep neural network inference
by Jean-François DAVID (LaBRI - Laboratoire Bordelais de Recherche en Informatique)
The defense will take place at 14h30 - Ada Lovelace Centre Inria de l'université de Bordeaux, 200 Av. de la Vieille Tour, 33405 Talence, FRANCE
in front of the jury composed of
- Olivier BEAUMONT - Directeur de recherche - Université de Bordeaux - Directeur de these
- Camille COTI - Professeure - École de Technologie Supérieure - Rapporteur
- Pierre MANNEBACK - Professeur émérite - Université de Mons - Rapporteur
- Akka ZEMMARI - Professeur des universités - Université de Bordeaux - Examinateur
The rise of deep neural networks has revolutionized many fields, notably computer vision and natural language processing. However, the growing size of these models imposes increasingly high computational requirements that challenge current computing infrastructures. Graphics Processing Units (GPUs), although optimized for massively parallel computations, can become overwhelmed by a high rate of inference requests, leading to congestion and escalating latency. At the same time, Central Processing Units (CPUs), which are often paired with GPUs for management tasks, remain underutilized despite their potential for handling some less computation-intensive tasks. It therefore seems appropriate to use available CPUs to assist GPUs in performing inference computations. In this thesis, we propose StarONNX, a solution aimed at accelerating inference computations on heterogeneous computing systems that combine CPUs and GPUs, particularly under heavy inference loads. StarONNX is based on the integration of partitions of DNN models obtained using the METIS partitioning tool, which allows for an asymmetric division of the computational workload. StarONNX builds upon StarPU, an execution system optimized for the dynamic scheduling of tasks on heterogeneous multicore architectures, and ONNX Runtime, a high-performance engine for inferring models in the ONNX format. Through this combination, StarONNX leverages the specific capabilities of each processor to improve both inference throughput and latency while maintaining efficient resource management. Our approach is based on combining methods from the high-performance computing domain to maximize the use of available resources. First, we take advantage of the possibility of overlapping computations and communications between processors over time. Second, we optimize the use of computing resources by grouping CPU cores to execute tasks. Third, we introduce pipelining of inference tasks, allowing different stages of processing to be executed simultaneously on GPUs and CPUs, thereby increasing both parallelism and throughput. When comparing our solution to NVIDIA's Triton Inference Server, we observe notable improvements in minimizing latency. Our solution pushes the system's congestion threshold to higher throughputs and fully exploits the capabilities of CPUs, which were previously underutilized. Despite an added latency overhead introduced by partitioning, this remains acceptable for partitioning into two or three segments.
ED Sciences Chimiques
Study of Non-Intentionally Added Substances (NIAS) in polypropylene and potential effects of recycling and UV aging
by Zakir AMIROV (Laboratoire de Chimie des Polymères Organiques)
The defense will take place at 9h30 - Amphi CRPP Paul Pascal Research Center (CRPP) 115 Avenue du Dr Albert Schweitzer, 33600 Pessac
in front of the jury composed of
- Véronique COMA - Associate Professor - Université de Bordeaux - Directeur de these
- Etienne FLEURY - Professeur - INSA (Lyon) - Rapporteur
- Emmanuel RICHAUD - Professeur des universités - PIMM (Paris) - Rapporteur
- Patrick NAVARD - Directeur de recherche émérite - CEMEF (Valbonne) - Examinateur
- Etienne GRAU - Assistant professor - LCPO (Pessac) - CoDirecteur de these
Non-Intentionally Added Substances (NIAS) are a major concern for regulatory and public health reasons, due to their potential impacts on health and the environment, particularly through their possible migration from packaging materials to food. To better understand and manage these dangers, this study investigated the impact of extrusion conditions, recycling, and aging on the formation of NIAS in a model polypropylene (PP). Initially, a model PP was formulated with various additives (Alkanox 240, BHT, UV326, UV944) to identify and quantify NIAS using multi-analytical techniques (GC, NMR). NIAS such as 2,4-di-tert-butylphenol, mono[2,4-bis(1,1-dimethylethyl)phenyl] ester, bis[2,4-bis(1,1-dimethylethyl)phenyl] ester, dihydroxyphosphine oxide, and tris(2,4-di-tert-butylphenyl) phosphate derived from Alkanox were identified and quantified. NIAS derived from BHT were also observed, although their low concentration did not allow for precise identification and quantification. In contrast, no NIAS derived from UV326 and UV944 were detected, although their presence combined with other additives, helped reduce the degradation of PP and Alkanox 240. The impact of extrusion (time, temperature, shear) on NIAS formation was then studied, with repeated cycles to simulate the recycling process. The process included three successive extrusion cycles at 240°C for 5 minutes, at a speed of 100 rpm. At each cycle, the additives degraded, leading to an increase in NIAS quantity. To limit PP degradation and simulate real industrial recycling conditions, additives were reintroduced at each stage. Despite this measure, a significant increase in the quantity and number of NIAS was observed. Finally, in the third phase, the impact of UV aging on the NIAS profile in PP was studied, considering the possibility of irradiation prior to recycling processes, such as during material usage. Additive degradation and NIAS formation were accelerated based on irradiation duration and the type of additive. In conclusion, this study provides key insights into the formation of NIAS throughout the lifecycle of plastics, thus contributing to the development of safer materials that comply with current regulations. This research can be directly applied to the identification and quantification of NIAS within the framework of quality control processes for PP materials, ensuring their safety and reliability prior to use.
ED Sociétés, Politique, Santé Publique
Evolution of the Framework for Health Economic Evaluation
by Sandrine BOURGUIGNON (Bordeaux Population Health Research Center)
The defense will take place at 13h00 - Salle Chastang à l'Isped Campus Carreire, 146 rue Léo Saignat, 33000 Bordeaux
in front of the jury composed of
- JEROME WITTWER - Professeur des universités - Université de Bordeaux - Directeur de these
- Hélène JACQMIN-GADDA - Directrice de recherche - UNIVERSITE DE BORDEAUX - Examinateur
- Isabelle BORGET - Professeure des universités - praticienne hospitalière - Service d'Etudes et Recherche en Economie de la Santé - Examinateur
- Antoine PARIENTE - Professeur des universités - Université de Bordeaux - Examinateur
- Florence JUSOT - Directrice de recherche - Université Paris Dauphine - Rapporteur
- Pascal PAUBEL - Professeur des universités - praticien hospitalier - Paris-Cité Université - Rapporteur
EVOLUTION OF THE HEALTHCARE ECONOMIC EVALUATION FRAMEWORK Health Economic evaluation has evolved in France under the expertise of the Haute Autorité de Santé (HAS) and the Economic and Public Health Commission called CEESP. Initially focused on the ICER by cost-utility study, economic evaluation plays a role in the prices negociation of healthcare products. However, these analyses remain limited, as they do not consider the broarder impacts on the healthcare system and society. Studies such as the evaluation of Mitraclip TM demonstrate the value of economic modelling but also highlight the lack of consideration of organizational impacts. The study of the reversal agent idaracizuma underscores the importance of structured expert consensus methods, such as the Delphi Panel when granular in-hospital data are lacking. Finally, the budget impact analysis of ferric carboxymaltose reveals potential cost savings and opens the door to considering organizational and environmental impacts. The evolution of the economic evaluation framework requires the integration of new criteria. Publics health, organizational, environmental and equity impacts are all factors that decision-makers could leverage to expand the scope of information supporting their decisions. These criteria could be integrated into existing health economic modelling methods. The structure of data is essential to ensure access to robust, comprehensive datasets. Sufficient granularity is mandatory to explore population health approaches considering geographical and socio-economic specificities. The inclusion of these new criteria would broaden the concept of value in healthcare. A more understandable and multidimensional approach would enable better resource allocation while ensuring equitable access to innovations. This works advocates for an evolution o health economic evaluation to incorporate a more global and pragmatic vision of the value of healthcare products. Such an approach is essential for ensuring the sustainability of healthcare systems and moving beyond the narrow focus on drug prices.