Self-aware memory: managing distributed memory in an autonomous multi-master environment

A major problem considering parallel computing is maintaining memory consistency and coherency, and ensuring ownership and access rights. These problems mainly arise from the fact that memory in parallel and distributed systems is still managed locally, e.g. using a combination of shared-bus- and directory-based approaches. As a result, such setups do not scale well with system size and are especially unsuitable for systems where such centralized management instances cannot or must not be employed. As a potential solution to this problem we present SaM, the Self-aware Memory architecture. By using self-awareness, our approach provides a novel memory architecture concept targeting multimaster systems with special focus on autonomic, self-managing systems. Unlike previous attempts, the approach delivers a holistic, yet scalable and cost-friendly solution to several memory-related problems including maintaining coherency, consistency, and access rights.