Performance demands of emerging domains such as artificial intelligence, machine learning and vision, Internet-of-things etc., continue to grow. Meeting such requirements on modern multi/many core systems with higher power densities, fixed power and energy budgets, and thermal constraints exacerbates the run-time management challenge. This leaves an open problem on extracting the required performance within the power and energy limits, while also ensuring thermal safety. Existing architectural solutions including asymmetric and heterogeneous cores and custom acceleration improve performance-per-watt in specific design time and static scenarios. However, satisfying applications’ performance requirements under dynamic and unknown workload scenarios subject to varying system dynamics of power, temperature and energy requires intelligent run-time management. 

Adaptive strategies are necessary for maximizing resource efficiency, considering i) diverse requirements and characteristics of concurrent applications, ii) dynamic workload variation, iii) core-level heterogeneity and iv) power, thermal and energy constraints. This dissertation proposes such adaptive techniques for efficient run-time resource management to maximize performance within fixed budgets under unknown and dynamic workload scenarios. Resource management strategies proposed in this dissertation comprehensively consider application and workload characteristics and variable effect of power actuation on performance for pro-active and appropriate allocation decisions. Specific contributions include i) run-time mapping approach to improve power budgets for higher throughput, ii) thermal aware performance boosting for efficient utilization of power budget and higher performance, iii) approximation as a run-time knob exploiting accuracy performance trade-offs for maximizing performance under power caps at minimal loss of accuracy and iv) co-ordinated approximation for heterogeneous systems 
through joint actuation of dynamic approximation and power knobs for performance guarantees with minimal power consumption. 

The approaches presented in this dissertation focus on adapting existing mapping techniques, performance boosting strategies, software and dynamic approximations to meet the performance requirements, simultaneously considering system constraints. The proposed strategies are compared against relevant state-of-the-art run-time management frameworks to qualitatively evaluate their efficacy.