The proposed research investigates collaborative coherence, a novel approach for building high-performance shared-memory multiprocessors. In this approach local cache memories collaborate to minimize the occurrence and detrimental performance effects of costly off-chip communication due to permission and data requests. Several problematic behaviors in conventional multiprocessors can be avoided or mitigated by exploiting global knowledge about sharing and reference patterns. This research will result in a deeper understanding of the interactions between processors in shared-memory multiprocessors, and will lead to system designs that have dramatically higher performance and lower power consumption, but are simple, intuitive, and easy to design.
The research will have a broad impact on the computer architecture community as a whole, by its development of representative commercial workloads, realization of state-of-the-art simulation infrastructure, and invention of powerful and useful evaluation methodologies. The techniques and methods created will be applied towards the development of increasingly powerful and ubiquitous computing devices whose utility will have a substantial impact on the productivity and creativity of countless end users. 2006
