Expressing parallel computations over complex shared-memory data structures has always been a vexing issue in parallel programming. On one hand, popular task-based programming models do not provide first-class abstractions for isolation and locality. On the other, Actor-based programming naturally captures locality but is unsuitable for computations on large shared data structures. The present project partially bridges the gap between these two styles of parallelism through Chorus, a new programming model for parallel computations over unstructured, continually changing shared-memory data structures.
The key abstraction of Chorus is an object assembly: a local, isolated region in the heap equipped with a thread of control. Assemblies can imperatively modify themselves, merge with other assemblies, and split into smaller assemblies?through these operations over assemblies, Chorus captures unpredictable, dynamic changes to parallelism. This makes Chorus an ideal programming model for many irregular data-parallel applications (e.g., meshing, clustering), which exhibit fine-grained data-parallelism in typical executions but no parallelism in the worst case, and whose parallelization remains an open and difficult challenge.
The predicted outcomes of the project include new insights into the semantic foundations of Chorus and new language constructs integrating Chorus with existing abstractions for asynchronous task creation, directed synchronization, and locality. On the system-building end, the project will integrate Chorus with the Habanero Java parallel programming language, and implement a compiler and runtime for the resultant language. The performance and programmability of this language will be thoroughly evaluated using benchmarks largely consisting of emerging irregular workloads.
