Energy-efficient management of reconfigurable computers

Power and energy consumption are first-order constraints on the design and operation of computer systems today. Improving energy efficiency reduces the amount of energy needed to perform a given computation as well as enables more computation to be performed for the same amount of energy. This saves operational costs to use these systems as well as capital costs to provision for them.;Conventionally, energy proportionality (energy consumption in proportion to the work done) as proposed by Barroso and H{"{o}}lzle has been the gold standard of an ideal system's energy efficiency. While this model is valid for fixed-resource systems, modern systems are reconfigurable in many aspects, allowing them to adapt to changing workload characteristics. We show that reconfigurability invalidates the conventional notions of ideal energy proportionality if the system behaves super-proportionally providing more performance (or work) in proportion to the power (or energy) used. We propose a new ideal model, Energy Optimal Proportional (EOP), that subsumes the conventional model and improves upon it by also accounting for super-proportional systems.;EOP can guide system designers to improve the maximum efficiency attainable over the operating range and forms a basis for comparisons of energy efficiency across systems. Power-performance Pareto optimality, on the other hand, can guide system operators to manage load and configure resources appropriately to make the current system execute efficiently. We propose a new intellectual framework that interrelates these two complementary energy efficiency goals.;The rest of this dissertation focuses on energy-efficient management. We develop new reactive governors that coordinate processor frequency (and voltage) and hardware prefetching to improve energy efficiency on a real (Haswell) server. We also propose a space-efficient hardware mechanism to estimate temporal locality (reuse) in cache accesses. The estimated distributions can be used by our new analytical models for cache performance to drive resizing decisions of the last-level cache.;Finally, we propose a new semantics-based classification system for system reconfiguration capabilities. We hope that this classification will be insightful to future researchers while exploring the space of reconfigurable systems, in categorizing existing work and in identifying coordination options that have been less well explored.