Energy modelling of multi-threaded, multi-core software for embedded systems

Efforts to reduce energy consumption are being made across all disciplines. ICT’s contribution to global energy consumption and by-products such as CO2 emissions continues to grow, making it an increasingly significant area in which improvements must be made. This thesis focuses on software as a means to reducing energy consumption. It presents methods for profiling and modelling a multi-threaded, multi-core embedded processor at the instruction set level, establishing links between the software and the energy consumed by the underlying hardware. A framework is presented that profiles the energy consumption characteristics of a multi-threaded processor core, associating energy consumption with the instruction set and parallelism present in a multi-threaded program. This profiling data is used to build a model of the processor that allows instruction set simulations to be used to estimate the energy that programs will consume, with an average of 2.67 % error. The profiling and modelling is then raised to the multi-core level, examining a channel based message passing system formed of a network of embedded multi-threaded processors. Additional profiling is presented that determines network communication costs as well as giving consideration towards system level properties such as power supply efficiency. Then, this is used to form a system level energy model that can estimate consumption using simulations of multi-core programs. The system level model combines multiple instances of a core energy model with a network level communication cost model. The broader implications of this work are explored in the context of other embedded and multi-core processor architectures, identifying opportunities for expanding or transferring the models. The models in this thesis are formed at the instruction set level, but have been demonstrated to be effective at higher-levels of abstraction than instruction set simulation, through their support of further work carried out externally. This work is enabled by several pieces of development effort, including a profiling framework for taking power measurements of the devices under investigation, tools for programming, routing and debugging software on a multi-core hardware platform called Swallow, and enhancements to an instruction set simulator for the simulation of this multi-core system. Through the work of this thesis, an embedded software developer for multi-threaded and multi-core systems is equipped with tools, techniques and new understanding that can help them in determining how their software consumes energy. This raises the status of energy efficiency in the software development cycle as we continue our efforts to minimise the energy impact of the world’s embedded devices.