The use of multiprocessor tasks (M-tasks) has been shown to be successful for mixed task and data parallel implementations of algorithms from scientific computing. The approach often leads to an increase of scalability compared to a pure data parallel implementation, but restricts the data exchange between M-tasks to the beginning or the end of their execution, expressing data or control dependencies between M-tasks.
In this article, we propose an extension of the M-task model to communicating M-tasks (CM-tasks) which allows communication between M-tasks during their execution. In particular, we present and discuss the CM-task programming model, programming support for designing CM-task programs, and experimental results. Internally, a CM-task comprises communication and computation phases. The communication between different CM-tasks can exploit optimized communication patterns for the data exchange between CM-tasks, e.g., by using orthogonal realizations of the communication. This can be used to further increase the scalability of many applications, including time-stepping methods which use a similar task structure for each time step. This is demonstrated for solution methods for ordinary differential equations.