ECHO · Extending Coherence for Hardware-Driven Optimizations in Multicore Architectures

Multicore processors are present nowadays in most digital devices, from smartphones to high-performanceservers. The increasing computational power of these processors is essential for enabling many importantemerging application domains such as big-data, media, medical, or scientific modeling. A fundamentaltechnique to improve performance is speculation, a technique that consists in executing work before it isknown if it is actually needed. In hardware, speculation significantly increases energy consumption byperforming unnecessary operations, while speculation in software (e.g., compilers) is not the default thuspreventing performance optimizations. Since performance in current multicores is limited by their powerbudget, it is imperative to make multicores as energy-efficient as possible to increase performance evenfurther.In a multicore architecture, the cache coherence protocol is an essential component since its unique butchallenging role is to offer a simple and unified view of the memory hierarchy. This project envisions thatextending the role of the coherence protocol to simplify other system components will be the key toovercome the performance and energy limitations of current multicores. In particular, ECHO proposes toadd simple but effective extensions to the cache coherence protocol in order to (i) reduce and eveneliminate misspeculations at the processing cores and synchronization mechanisms and to (ii) enablespeculative optimizations at compile time. The goal of this innovative approach is to improve theperformance and energy efficiency of future multicore architectures. To accomplish the objectivesproposed in this project, I will build on my 14 years expertise in cache coherence, documented in over 40publications of high impact.