Foundations of Weakly Consistent Objects for Shared-Memory Programming

Despite being intuitive, strong consistency necessitates frequent global coordination, significantly limiting the scalability of concurrent systems. In shared-memory programming, the need to relax the consistency guarantees has long been recognized, spurring extensive research and driving the adoption of weak memory models in programming languages. However, the situation is markedly different at the level of concurrent objects. For objects---designed to provide common functionalities to clients without requiring knowledge of the internal implementation---linearizability remains the de facto standard, leaving programmers with limited means to use formally specified objects free from the performance penalties of strong consistency.Our overarching goal is to address this gap by designing foundational weakenings of linearizability for general-purpose programming with various forms of weakly consistent memory, while retaining the benefits of linearizability. In particular, we will develop correctness notions that enable client reasoning through in-language observational refinement, analogous to how linearizability readily corresponds to refinement w.r.t. atomic reference implementations. This correspondence will also be the key to achieving compositionality of different correctness criteria within a single program---an essential yet largely overlooked necessity in software development. Additionally, we will develop theoretical foundations that establish impossibility results and lower bounds on coordination costs for various objects under different correctness criteria.To achieve our goal, we will build on recent advancements by the PI in the study of weak memory models. We envision our outcomes fostering the principled development and use of libraries under weak memory consistency, driving significant improvements in performance and scalability, and positioning us for the emerging era of distributed memory powered by CXL infrastructures.