Designing memory consistency models for shared-memory multiprocessors

TL;DR: This thesis establishes a unifying framework for designing memory models that can adequately satisfy the 3P criteria and applies debugging techniques for sequential consistency to two of the SCNF models to alleviate the problem of debugging programs onSCNF models.

Abstract: The memory consistency model (or memory model) of a shared-memory multiprocessor system influences the performance and the programmability of the system. The most intuitive model for programmers, sequential consistency, restricts many performance-enhancing optimizations. For higher performance, several alternative models have been proposed. The hardware-centric nature of these models, however, makes them difficult to program and inhibits portability. We use a 3P criteria of programmability, portability, and performance to assess memory models, and find current models lacking. This thesis establishes a unifying framework for designing memory models that can adequately satisfy the 3P criteria. The first contribution of this thesis is a programmer-centric methodology, called sequential consistency normal form (SCNF), for specifying memory models. SCNF is based on the observation that a system can employ performance-enhancing optimizations without violating sequential consistency if the system has some information about the program. An SCNF model is a contract between the system and the programmer, where the system guarantees high performance and sequential consistency only if the programmer provides certain information about the program. Insufficient information gives lower performance, but incorrect information violates sequential consistency. SCNF satisfies the 3P criteria of programmability (by providing sequential consistency), portability (by providing a common interface of sequential consistency across all models), and performance (by only requiring sequential consistency for programs with correct information). The second contribution demonstrates the effectiveness of SCNF by applying it to optimizations of previous hardware-centric models, resulting in four SCNF models. Although based on intuition similar to the hardware-centric models, these SCNF models are easier to program, enhance portability, and allow more implementations (with potentially higher performance) than the corresponding hardware-centric models. The third contribution culminates the above work by exposing a large part of the design space of SCNF models. SCNF models are difficult to design because the relationship between system optimizations and programmer information is complex. We simplify this relationship and use it to characterize and explore the design space. The final contribution concerns debugging programs on SCNF models. While debugging, the programmer may unknowingly provide incorrect information, violating sequential consistency. We apply debugging techniques for sequential consistency to two of the SCNF models to alleviate this problem.