Improving static scheduling using inter-task concurrency measures

A fundamental issue affecting the performance of parallel applications running on distributed systems is the assignment of tasks to processors. This paper shows the effectiveness in scheduling strategies derived from the use of the temporal behaviour of tasks included in the new TTIG (Temporal Task Interaction Graph) model. Experimentation was performed for a set of C+PVM applications running in a PVM platform. These applications were on the one hand synthetic programs whose communication topology matches certain well-known regular graph families such as trees, pipes and meshes and programs with irregular communication patterns. Additionally, a real image processing application was modelled and executed. The TTIG model has been shown to be effective in all cases compared with the classical TIG (Task Interaction Graph) and with the PVM default allocation scheme, and facilitates the development of new more efficient scheduling strategies.