The Tuning Language for Concurrent Collections

Concurrent Collections (CnC) is a programming model for parallel systems. A novel aspect of this model is that there is a clear separation of the specification of the application semantics, called the domain specification, and the tuning specification, which maps the domain spec to a platform. The domain spec is declarative specification that indicates the constraints on execution. These correspond to data dependences and control dependences. The domain spec can be written by a domain expert who does not necessarily have knowledge of the target architecture. The separation of concerns isolates the domain expert from the tuning facility. This isolation permits the tuning language to provide strong capabilities for control and flexibility for a tuning expert. In addition, the separation means that a given domain spec can have multiple tuning specs. There might be multiple targets, and within the same target there might be distinct goals, such as performance vs. power. This paper introduces the CnC language for tuning specifications in some detail and show examples of its use. The goal of the tuning language is to provide capability for mapping the parallelism implicit in the domain spec for high-performance execution on a target platform. The focus of this activity is locality. The basic concept in the tuning language is the affinity group, a set of computations that the tuner suggests executing close in time and space. Hierarchical affinity groups allow the specification of relative levels of affinity. They provide a mechanism that allows the programmer to specify locality, while allowing but not requiring him to distinguish between spatial and temporal locality. We will use Cholesky factorization to show the use of multiple tuning specs for a single domain spec. The tuning spec is under active design.