No abstract

No abstract

Abstract. We present a non-operational approach to specifying and analyzing shared memory consistency models. The method uses higher order logic to capture a complete set of ordering constraints on execution traces, in an axiomatic style. A direct encoding of the semantics with a constraint logic programming language provides an interactive and incremental framework for exercising and verifying finite test programs. The framework has also been adapted to generate equivalent boolean satisfiability (SAT) problems. These techniques make a memory model specification executable, a powerful feature lacked in most non-operational methods. As an example, we provide a concise formalization of the Intel Itanium memory model and show how constraint solving and SAT solving can be effectively applied for computer aided analysis. Encouraging initial results demonstrate the scalability for complex industrial designs.

SUMMARYGiven the complicated nature of modern shared memory systems, it is vital to have a systematic approach to specifying and analyzing memory consistency requirements. In this paper, we present the UMM specification framework, which integrates two key features to support memory model verification: (i) it employs a simple and generic memory abstraction that can capture a large collection of memory models as guarded commands with a uniform notation, and (ii) it provides built-in model checking capability to enable formal reasoning about thread behaviors. Using this framework, memory models can be specified in a parameterized style-designers can simply redefine a few bypassing rules and visibility ordering rules to obtain an executable specification of another memory model. We formalize several classical memory models, including Sequential Consistency, Coherence, and PRAM, to illustrate the general techniques of applying this framework. We then provide an alternative specification of the Java memory model, based on a proposal from Manson and Pugh, and demonstrate how to analyze Java thread semantics using model checking. We also compare our operational specification style with axiomatic specification styles and explore a mechanism that converts a memory model definition from one style to the other.

No abstract