Local Qualification Inference for Titanium

Titanium is a parallel programming language that presents a single global address space for all concurrent computations. To improve performance on distributed memory architectures, programmers may explicitly qualify selected references as local to a single process. This paper proposes a static inference system for automatically applying “local” qualifications wherever possible, within the constraints of the Titanium type system. A partial prototype implementation of the analysis has been completed with the help of the Bane analysis toolkit. We describe and evaluate the analysis itself, as well as the implementation strategy by which the analysis has been deployed.