Event-based control for coordinated thermal, power and performance management in high-density microprocessors

This thesis is part of a research work aimed at defining a new control strategy which integrates the thermal and the power-performance management in multicore processors. The intention is to develop a controller that is able to manage the current multi-core architectures and especially the future ones, in which modern technologies are involved in order to provide an higher power density in reduced volumes. Special attention, as future application, is given to the 3D-stacking architectures which exhibit a larger power density but, differently from the planar geometries, not all their layers are directly connected to the heat sink and this increases the risk of thermal runaway. Furthermore, in such context, it appears essential to develop an integrated controller that is able to take action on the power-performance optimization and also to ensure a suitable thermal control, in order to provide reliability of the overall system and energy saving. Nowadays, the control strategies used in commercial processors divide the power-performance and the thermal issues. The former is handled at the software level by the operating system, as the Linux governors do, while the latter is handled in hardware in the form of a thermal limit, not integrated in the overall optimization. Up to now, under these considerations, the solution adopted in industry has allowed to reach only sub-optimal results; the objective of this thesis is instead to unify the thermal and the power-performance management in a single controller that is able to administrate synergically power, performance and thermal issues. Hence, the strategy adopted in this research work involves the implementation of a thermal controller, whose policy is based on the event-based theory, with the objective to generate a control action that solves both the thermal and power-performance management; this is achievable thanks to the introduction of a connection between the hardware and the software level of the machine that allows to split the control action between them. Moreover, this thesis is focused on the validation of the event-based control policy implemented in related research works. Many experiments have been performed on a test machine built \ad hoc" in laboratory with the aim to compare the behaviour of our controller with those available on the market, especially the thermal controller provided by the Intel (thermal daemon) and that can be found on all the latest Intel processors. These tests highlight strengths and weaknesses of our controller based on the event-based theory even running several benchmarks under different control requirements. Another goal of this thesis has been the improvement of the event-based control policy itself by implementing an adaptive event generation policy to enhance thermal performances. Its development has been based on the definition of a suitable temperature threshold collection and on the definition of a timeout one, in order to minimize the number of events and to avoid all the useless ones. The resulting event-based controller with these improvements was shown to provide a lower computational cost and an higher performance level.