Memory and communication driven spatio-temporal scheduling on MPSoCs

Scheduling and executing software efficiently on contemporary embedded systems, featuring heterogeneous multi-processors, multiple power modes, complex memory hierarchies and advanced interconnects, is a daunting task. State-of-the-art tools that schedule software tasks to hardware resources face limitations: (1) either they do not take into account the interdependancies among processing, memory and communication constraints (2) or they decouple the problem of spatial assignment from temporal scheduling. As a result existing tools make sub-optimal spatio-temporal scheduling decisions. This paper presents a technique to find globally optimized solutions by co-exploring spatio-temporal schedules for computation, data storage and communication simultaneously, considering the interdependencies between them. Experiments on mapping exploration of an image processing application on a heterogeneous MPSoC platform show that this co-exploration methodology finds schedules that are more energy efficient, when compared to decoupled exploration techniques for the particular application and target platform.