A graph-oriented task manager for small multiprocessor systems

A task manager that dynamically decodes the data-dependent task graph is a key component of general multiprocessor systems. The emergence of small-scale parallel systems for multimedia and general-purpose applications requires the extraction of complex parallelism patterns. The small system size also allows the centralization of the task generation and synchronization. This paper proposes such a task manager. It uses a structured representation of the task dependence graph to issue and synchronize tasks. We describe several optimizations to extract more parallelism, discuss the software/hardware implementation issue and show it produces efficient parallelism exploitation in case of applications with complex parallelism patterns