Cette thèse propose un large éventail d'approches afin d'améliorer différents aspects de l'analyse des flux de puissance avec comme fils conducteur l'utilisation du processeurs graphiques (GPU). Si les GPU ont rapidement prouvés leurs efficacités sur des applications régulières pour lesquelles le parallélisme de données était facilement exploitable, il en est tout autrement pour les applications dites irrégulières. Ceci est précisément le cas de la plupart des algorithmes d'analyse de flux de puissance. Pour ce travail, nous nous inscrivons dans cette problématique d'optimisation de l'analyse de flux de puissance à l'aide de coprocesseur de type GPU. L'intérêt est double. Il étend le domaine d'application des GPU à une nouvelle classe de problème et/ou d'algorithme en proposant des solutions originales. Il permet aussi à l'analyse des flux de puissance de rester pertinent dans un contexte de changements continus dans les systèmes énergétiques, et ainsi d'en faciliter leur évolution. Nos principales contributions liées à la programmation sur GPU sont: (i) l'analyse des différentes méthodes de parcours d'arbre pour apporter une réponse au problème de la régularité par rapport à l'équilibrage de charge ; (ii) l'analyse de l'impact du format de représentation sur la performance des implémentations d'arithmétique floue. Nos contributions à l'analyse des flux de puissance sont les suivantes: (ii) une nouvelle méthode pour l'évaluation de l'incertitude dans l'analyse des flux de puissance ; (ii) une nouvelle méthode de point fixe pour l'analyse des flux de puissance, problème que l'on qualifie d'intrinsèquement parallèle. / This thesis addresses the utilization of Graphics Processing Units (GPUs) for improving the Power Flow (PF) analysis of modern power systems. Currently, GPUs are challenged by applications exhibiting an irregular computational pattern, as is the case of most known methods for PF analysis. At the same time, the PF analysis needs to be improved in order to cope with new requirements of efficiency and accuracy coming from the Smart Grid concept. The relevance of GPU-enhanced PF analysis is twofold. On one hand, it expands the application domain of GPU to a new class of problems. On the other hand, it consistently increases the computational capacity available for power system operation and design. The present work attempts to achieve that in two complementary ways: (i) by developing novel GPU programming strategies for available PF algorithms, and (ii) by proposing novel PF analysis methods that can exploit the numerous features present in GPU architectures. Specific contributions on GPU computing include: (i) a comparison of two programming paradigms, namely regularity and load-balancing, for implementing the so-called treefix operations; (ii) a study of the impact of the representation format over performance and accuracy, for fuzzy interval algebraic operations; and (iii) the utilization of architecture-specific design, as a novel strategy to improve performance scalability of applications. Contributions on PF analysis include: (i) the design and evaluation of a novel method for the uncertainty assessment, based on the fuzzy interval approach; and (ii) the development of an intrinsically parallel method for PF analysis, which is not affected by the Amdahl's law.

Indiana University-Purdue University Indianapolis (IUPUI) / Current mainstream computing systems have become increasingly complex. Most of which have Central Processing Units (CPUs) that invoke multiple threads for their computing tasks. The growing issue with these systems is resource contention and with resource contention comes the risk of encountering a deadlock status in the system. Various software and hardware approaches exist that implement deadlock detection/avoidance techniques; however, they lack either the speed or problem size capability needed for real-time systems. The research conducted for this thesis aims to resolve issues present in past approaches by converging the two platforms (software and hardware) by means of the Graphics Processing Unit (GPU). Presented in this thesis are two GPU-based deadlock detection algorithms and one GPU-based deadlock avoidance algorithm. These GPU-based algorithms are: (i) GPU-OSDDA: A GPU-based Single Unit Resource Deadlock Detection Algorithm, (ii) GPU-LMDDA: A GPU-based Multi-Unit Resource Deadlock Detection Algorithm, and (iii) GPU-PBA: A GPU-based Deadlock Avoidance Algorithm. Both GPU-OSDDA and GPU-LMDDA utilize the Resource Allocation Graph (RAG) to represent resource allocation status in the system. However, the RAG is represented using integer-length bit-vectors. The advantages brought forth by this approach are plenty: (i) less memory required for algorithm matrices, (ii) 32 computations performed per instruction (in most cases), and (iii) allows our algorithms to handle large numbers of processes and resources. The deadlock detection algorithms also require minimal interaction with the CPU by implementing matrix storage and algorithm computations on the GPU, thus providing an interactive service type of behavior. As a result of this approach, both algorithms were able to achieve speedups over two orders of magnitude higher than their serial CPU implementations (3.17-317.42x for GPU-OSDDA and 37.17-812.50x for GPU-LMDDA). Lastly, GPU-PBA is the first parallel deadlock avoidance algorithm implemented on the GPU. While it does not achieve two orders of magnitude speedup over its CPU implementation, it does provide a platform for future deadlock avoidance research for the GPU.

Indiana University-Purdue University Indianapolis (IUPUI) / Adaptive-optics optical coherence tomography (AOOCT) is a technology that has been rapidly advancing in recent years and offers amazing capabilities in scanning the human eye in vivo. In order to bring the ultra-high resolution capabilities to clinical use, however, newer technology needs to be used in the image reconstruction process. General purpose computation on graphics processing units is one such way that this computationally intensive reconstruction can be performed in a desktop computer in real-time. This work shows the process of AOOCT image reconstruction, the basics of how to use NVIDIA's CUDA to write parallel code, and a new AOOCT image reconstruction technology implemented using NVIDIA's CUDA. The results of this work demonstrate that image reconstruction can be done in real-time with high accuracy using a GPU.

Indiana University-Purdue University Indianapolis (IUPUI) / Adaptive optics-optical coherence tomography (AO-OCT) is a non-invasive method of imaging the human retina in vivo. It can be used to visualize microscopic structures, making it incredibly useful for the early detection and diagnosis of retinal disease. The research group at Indiana University has a novel multi-camera AO-OCT system capable of 1 MHz acquisition rates. Until this point, a method has not existed to process data from such a novel system quickly and accurately enough on a CPU, a GPU, or one that can scale to multiple GPUs automatically in an efficient manner. This is a barrier to using a MHz AO-OCT system in a clinical environment. A novel approach to processing AO-OCT data from the unique multi-camera optics system is tested on multiple graphics processing units (GPUs) in parallel with one, two, and four camera combinations. The design and results demonstrate a scalable, reusable, extensible method of computing AO-OCT output. This approach can either achieve real time results with an AO-OCT system capable of 1 MHz acquisition rates or be scaled to a higher accuracy mode with a fast Fourier transform of 16,384 complex values.

Graphics Processing Units (GPUs) are emerging as the most popular accelerator for many applications, powering the core of Machine Learning applications and many computing-intensive workloads. GPUs have typically been consideredas accelerators, with Central Processing Units (CPUs) in charge of the mainapplication logic, data movement, and network connectivity. In these architectures,input and output data of network-based GPU-accelerated application typically traverse the CPU, and the Operating System network stack multiple times, getting copied across the system main memory. These increase application latency and require expensive CPU cycles, reducing the power efficiency of systems, and increasing the overall response times. These inefficiencies become of higher importance in latency-bounded deployments, or with high throughput, where copy times could easily inflate the response time of modern GPUs. The main contribution of this dissertation is towards a GPU-centric network architecture, allowing GPUs to initiate network transfers without the intervention of CPUs. We focus on commodity hardware, using NVIDIA GPUs and Remote Direct Memory Access over Converged Ethernet (RoCE) to realize this architecture, removing the need of highly homogeneous clusters and ad-hoc designed network architecture, as it is required by many other similar approaches. By porting some rdma-core posting routines to GPU runtime, we can saturate a 100-Gbps link without any CPU cycle, reducing the overall system response time, while increasing the power efficiency and improving the application throughput.The second contribution concerns the analysis of Clockwork, a State-of-The-Art inference serving system, showing the limitations imposed by controller-centric, CPU-mediated architectures. We then propose an alternative architecture to this system based on an RDMA transport, and we study some performance gains that such a system would introduce. An integral component of an inference system is to account and track user flows,and distribute them across multiple worker nodes. Our third contribution aims to understand the challenges of Connection Tracking applications running at 100Gbps, in the context of a Stateful Load Balancer running on commodity hardware. / <p>QC 20240315</p>

With the advent of low-energy buildings, the need for accurate building performance simulations has significantly increased. However, for the time being, the thermo-aeraulic effects are often taken into account through simplified or even empirical models, which fail to provide the expected accuracy. Resorting to computational fluid dynamics seems therefore unavoidable, but the required computational effort is in general prohibitive. The joint use of innovative approaches such as the lattice Boltzmann method (LBM) and massively parallel computing devices such as graphics processing units (GPUs) could help to overcome these limits. The present research work is devoted to explore the potential of such a strategy. The lattice Boltzmann method, which is based on a discretised version of the Boltzmann equation, is an explicit approach offering numerous attractive features: accuracy, stability, ability to handle complex geometries, etc. It is therefore an interesting alternative to the direct solving of the Navier-Stokes equations using classic numerical analysis. From an algorithmic standpoint, the LBM is well-suited for parallel implementations. The use of graphics processors to perform general purpose computations is increasingly widespread in high performance computing. These massively parallel circuits provide up to now unrivalled performance at a rather moderate cost. Yet, due to numerous hardware induced constraints, GPU programming is quite complex and the possible benefits in performance depend strongly on the algorithmic nature of the targeted application. For LBM, GPU implementations currently provide performance two orders of magnitude higher than a weakly optimised sequential CPU implementation. The present thesis consists of a collection of nine articles published in international journals and proceedings of international conferences (the last one being under review). These contributions address the issues related to single-GPU implementations of the LBM and the optimisation of memory accesses, as well as multi-GPU implementations and the modelling of inter-GPU and internode communication. In addition, we outline several extensions to the LBM, which appear essential to perform actual building thermo-aeraulic simulations. The test cases we used to validate our codes account for the strong potential of GPU LBM solvers in practice.