Our energy optimisation methods and solutions including dynamic voltage and frequency scaling (DVFS) help achieve more peak computational power with less energy consumption.
Mobile devices, ubiquitous computing and smart grids include computing devices that are often battery-powered or their energy consumption is significant for overall feasibility, including environmental sustainability. DVFS and other optimisation methods offer a dynamic and flexible way of controlling the amount of energy used to perform necessary computations. These methods allow various operational profiles for either maximising energy preservation or offering more peak load processing capacity.
Operating system-level performance monitoring and adaptive scaling for energy optimisation
We have successfully implemented performance monitoring and prediction at the operating system level, where relevant information is available for scaling decisions more than at HW/instruction level. The up- and down-scaling of performance can use context information of the computing load, for example by prioritising certain tasks with higher real-time requirements. Using such information allows for a quicker reaction to sudden load peaks of high importance than traditional scaling algorithms based only on the total amount of recent load history.
For example, a mobile handheld device may have both static and varying computing load producing processes running simultaneously. Additional variation comes from the amount of user interaction, which tends to create unpredictable load peaks. The scaling decisions can treat these different types of load with different heuristics and guide the processor to react to them differently, thus optimising the balance between energy minimisation and performance needs.
We have developed new algorithms, such as PROPHET, for optimising overall energy consumption with tolerable performance degradation in user-interactive and thus possibly unpredictable scenarios. We have also analysed and tuned the parameters of various algorithms for user-interactive application use cases to fulfil responsiveness requirements, by keeping the user interface delays under the limit of what human users still perceive as immediate.
Our methods for energy optimisation have been applied in the telecoms business area within commercial-level product prototypes. These methods are, however, applicable to any domain where energy consumption minimisation is important while still offering good maximum performance of processors and system-on-chips.
We currently possess one patent for "Characterising run-time properties of computing system" (US20090006456 A1) and we have written scientific publications about DVFS and load prediction algorithms, such as "PROPHET: speculative load prediction algorithm for dynamic performance scaling" (IWCMC 2007) and "Optimising dynamic performance scaling for user interface performance" (5th International Conference on Mobile Technology, Applications and Systems, Mobility Conference 2008). In addition, we have several publications about managing the non-functional requirements of software-intensive embedded systems. We have also contributed to patent applications with our industrial customers with regard to performance scaling and energy optimisation.