Power optimisation for multi-processor dataflow applications - Jaco van de Pol

Synchronous Data Flow (SDF) provides a convenient formalism for the design of streaming applications on embedded multi-processor systems. SDF tools provide efficient algorithms for mapping tasks onto processors, and scheduling them in time.

In reality, hardware platforms are heterogeneous. Processors provide different capabilities and vary in speed, memory, and energy consumption. The design problem is complicated by constraints on the mapping of tasks to processors. Moreover, a tradeoff is required between time and energy consumption.

We propose to complement a functional SDF description by a realistic Processor Application Model (PAM). The PAM describes constraints, speed and power consumption of the processors. It also describes the dynamic power management capabilities of the hardware system. Examples include voltage and frequency scaling, which can even be applied to individual voltage-frequency islands.

We provide a systematic and automatic translation of the application (SDF) and the hardware (PAM) to Timed Automata with costs. Subsequently, we use UPPAAL to determine efficient mappings and schedules on a minimal set of processors, and UPPAAL CORA to compute optimal schedules and energy-management strategies, taking into account trade-offs between time and energy.

(this is joint work with Waheed Ahmad, Philip Hölzenspies and Mariëlle Stoelinga)