Hybrid coherence for scalable multicore architectures

This dissertation describes a cache architecture and memory model for 1000core microprocessors. Our approach exploits workload characteristics and programming model assumptions to build a hybrid memory model that incorporates features from both software-managed coherence schemes and hardware-managed cache coherence. The goal is to achieve the scalability found in compute accelerators, which support relaxed ordering of memory operations and programmermanaged coherence, while providing a programming interface that is akin to the strongly ordered cache coherent memory models found in general-purpose multicore processors today. The research presented in this dissertation supports the following thesis: To be scalable and programmable, future multicore systems require a cached, singleaddress space memory hierarchy. A hybrid software and hardware approach to coherence management is required to support such a memory hierarchy in 1000core processors and is achievable only by leveraging the characteristics of target applications and system software. We motivate a hybrid memory model and present our approach to addressing the challenges facing such a model. We discuss and evaluate a scalable 1024core architecture, workloads that we see as targets for such an architecture, a memory model that relies on software management of coherence, and scalable hardware coherence schemes. Using these components, we develop the software and hardware support for a hybrid memory model. We demonstrate that our techniques can be used to reduce hardware design complexity, to increase software scalability, or to combine the two.