Task Variants with Different Scratchpad Memory Consumption in Multi-Task Environments

We present an approach which schedules task sets using scratchpad memory (SPM ) in an embedded multi-task system with realtime constraints. A new task model is introduced, where each task is represented by di erent pre-compiled variants which di er in the amount of scratchpad memory used. A higher use of SPM leads to smaller runtimes of a task. Moreover, the energy consumption is reduced by replacing memory accesses by SPM accesses. Our heuristic method assembles a task set of these variants by choosing one variant per task. After selecting candidates from the pre-computed set of task variants, the task set can be handled by a real-time scheduler like EDF. Our approach is able to build a new incremental task set and feasible transition in dynamically changing environments. Furthermore we show an extension of our approach to multicore environments.