@PHDTHESIS{ 2023:60244047, title = {Enhancing lifetime reliability of manycore systems through reinforcement learning-based task management}, year = {2023}, url = "https://tede2.pucrs.br/tede2/handle/tede/11734", abstract = "This research tackles the challenge of improving the lifetime reliability of manycore systems, a critical issue in microelectronics. The current state-of-the-art in Dynamic Thermal Management (DTM) and Dynamic Reliability Management (DRM) techniques present the following gaps: (i) system underutilization in patterning approaches or adoption of complex heuristics; (ii) works focusing only on temperature (DTM) or reliability (DRM); (iii) proposals considering few aging effects. The primary goal of this Thesis is to address the issue of early degradation in manycore systems resulting from temperature-amplified wear-out effects, encompassing the development and execution of strategies to manage tasks in ways that mitigate these effects. The central claim of the Thesis is that task management based on reinforcement learning (RL) can enhance manycore systems lifetime reliability. The research adopts an innovative approach using an RL algorithm for task management. This method involves building models to predict system degradation and dynamically modifying task allocations to minimize long-term wear. The research employs simulations to verify the effectiveness of the developed models and algorithms. The significant contribution of this Thesis is the creation of the "Failure In Time-aware Learning Heuristic for Application Allocation" (FLEA), which manages temperature and reliability concomitantly. Results show that FLEA lowers the rate of system degradation compared to conventional task management approaches. The results data present an enhancement in system reliability and lifetime. FLEA represents an advancement in management, combining reinforcement learning techniques with task management strategies to proactively increase lifetime. This Thesis provides insights into the design and management of manycores. It paves the way for developing more sophisticated reinforcement learning models for systems management.", publisher = {Pontif?cia Universidade Cat?lica do Rio Grande do Sul}, scholl = {Programa de P?s-Gradua??o em Ci?ncia da Computa??o}, note = {Escola Polit?cnica} }