The efficiency of an application can have a huge effect on how long a device will run on its batteries. With no significant increases in battery capacity, yet applications requiring longer battery life, it falls to software to be more efficient with its use of energy. This is particularly salient in deeply embedded systems which may have to run on batteries for years. The high-level nature of writing a computer program detaches the programmer from the underlying hardware, allowing programmers to quickly write code that will be functional on a diverse set of devices. However, it also complicates writing efficient software, since the code may go through many transformations before it is executed, each of which can result in larger energy consumption without this being easily observable. Therefore, methods of increasing software energy efficiency are needed. Compiler optimisations provide an ideal way to achieve this, providing the ability to automatically transform software into a more efficient form. Typically, compilers have focused on making an application fast - there exist hundreds of optimisations to decrease runtime, and often these are effective at reducing energy. However, few optimisations exist to specifically decrease energy consumption. This thesis explores the differences between automated ways to reduce energy consumption and execution time at the compiler level. By exploring an extensive selection of existing compiler optimisations, through statistical techniques and genetic algorithms, it is discovered that current optimisations only reduce energy consumption because they reduce the execution time; while the optimisations affect the power dissipation of the device, this effect is incidental. The lack of optimisations which affect power dissipation implies a class of compiler optimisations has been missed from conventional compilers. This thesis develops two new optimisations belonging to this class. To create these optimisations, low-level hardware-specific energy characteristics are identified, focusing on embedded systems. The characteristics are rigorously modelled, allowing the compiler to make optimisation decisions at a higher level of abstraction. One of the optimisations manages to save up to 26% of the energy consumption, achieving this by focusing on average power reduction, rather than execution time reduction. The combination of the optimisation for energy with the existing optimisations for time is explored and found not to interact significantly with other optimisations, proving to be linearly composable. This again suggests that the energy optimisation belongs to a different class of compiler optimisation. To achieve further energy efficiency gains, a thorough vertical integration process needs to be introduced. This requires identifying (possibly) hardware-specific energy characteristics, modelling them at a level accessible to the compiler, and then making optimisation decisions based on these models. This level of integration is essential to bridge the levels of abstraction that exist between the hardware and software, and allows compilers to be successful at reducing applications' energy consumption and consequently increasing battery life.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:702470 |
| Date | January 2016 |
| Creators | Pallister, James |
| Contributors | Eder, Kerstin; Hollis, Simon J. |
| Publisher | University of Bristol |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
Page generated in 0.0131 seconds