MESSAGE-FLOW : A MODEL FOR THE CONCURRENT EXECUTION OF PROGRAMS ON MESSAGE-PASSING MULTICOMPUTERS

Muhammad  F. MUDAWWAR

The design of concurrent machines is based on execution models. The execution model defines the nature of an activity and its interaction with other concurrent activities. Most execution models use the concept of a process as the basis for an activity. Communication and synchronization of different parallel activities occur either through a common shared memory or through explicit message passing.

In the dataflow model, a simple operator that acts on its input data values is the basis of an activity. In essence, the dataflow model should exploit all the parallelism embedded in a program. In reality, it does not. Its weakest and most problematic feature is that it does not recognize the notion of state and storage. This can be a problem when modeling systems that change their behavior over time.

We propose a new execution model that brings the notion of state and storage to its focus. This model, called message flow, defines an activity as a message acting on a storage cell. Message flow graphs depict the flow of activities. Storage cells are represented by nodes and messages are represented by arcs. Large data structures can be distributed over storage cells. Computation is viewed as a message pattern on storage cells.