A Type System for the Derivation of Data Redistributions

Many algorithms can be formulated in a task or module oriented programming model where the data dependencies between subtasks restrict their execution order. We consider a two-level parallel computation model where task parallelism is expressed in a hierarchical module structure which abstracts from the internal behavior consisting of system or data parallel computations. The module language is extended with types that represent data distributions for distributed memory machines. We show how these types can be used for checking the consistency of data distributions and for the automatic derivation of redistributions when cooperating tasks use different distributions for the same data structure.