Abstract

Parallel programming languages that let multiple processors access shared data provide a variety of sharing mechanisms and memory models. Understanding a language's support for data sharing behavior is critical to understanding how the language can be used, and is also a component for numerous program analysis, optimization, and runtime clients. Languages that provide the illusion of a global address space, but are intended to work on machines with physically distributed memory, often distinguish between different kinds of pointers or shared data. The result can be subtle rules about what kinds of accesses are allowed in the application programs and implicit constraints on how the language may be implemented. This paper develops a basis for understanding the design space of these sharing formalisms, and codifies that understanding in a suite of type checking/inference systems that illustrate the trade-offs among various models.

Citations