From software to hardware: making dynamic multi-core processors practical
Micolet, Paul-Jules René Régis
Heterogeneous processors such as Arm’s big.LITTLE have become popular as they offer a choice between performance and energy efficiency. However, the core configurations are fixed at design time which offers a limited amount of adaptation. Dynamic Multi-core Processors (DMPs) bridge the gap between homogeneous and fully reconfigurable systems. They present a new way of improving single-threaded performance by running a thread on groups of cores (compositions) and with the ability of changing the processor topology on the fly, they can better adapt themselves to any task at hand. However, these potential performance improvements are made difficult due to two main challenges: the difficulty of determining a processor configuration that leads to the optimal performance and knowing how to tackle hardware bottlenecks that may impede the performance of composition. This thesis first demonstrates that ahead-of-time thread and core partitioning used to improve the performance of multi-threaded programs can be automated. This is done by analysing static code features to generate a machine-learning model that determines a processor configuration that leads to good performance for an application. The machine learning model is able to predict a configuration that is within 16% of the performance of the best configuration from the search space. This is followed by studying how dynamically adapting the size of a composition at runtime can be used to reduce energy consumption whilst maintaining the same speedup as the fastest static core composition. An analysis of how much energy can be saved by adapting the size of the composition at runtime is conducted, showing that dynamic reconfiguration can reduce energy consumption by 42% on average. A model is then built using linear regression which analyses the features of basic blocks being executed to determine if the current composition should be reconfigured; on average it reduces energy consumption by 37%. Finally the hardware mechanisms that drive core composition are explored. A new fetching mechanism for core composition is proposed, where cores fetch code in a round-robin fashion. The use of value prediction is also motivated, as large core compositions are more susceptible to data-dependencies. This new hardware setup shows massive potential. By exploring a perfect value predictor with perfect branch prediction and the new fetching scheme, the performance of a large core composition can be improved by a factor of up to 3x, and 1.88x on average. Furthermore, this thesis shows that state-of-the-art value prediction with a normal branch predictor still leads to good performance improvements, with an average of 1.33x to up to 2.7x speedup.