Evolutionary algorithms

Research areas: Cloud computing; Computer architectures; Cyber Physical Systems; Embedded systems; Evolutionary computation; Real time languages and systems

Background
Application-aware resource management in multiprocessor, high-performance and cloud computing
The mapping of application tasks onto computing platforms has critical impact over the performance and energy-efficiency of computer systems, as the mapping process partitions the application and allocates the platform resources that will execute each of the partitions. This problem has been addressed since the dawn of computing, but its importance has grown recently because of the additional complexity imposed by the potential availability of massive parallelism at the platform level. Many of the existing heuristics can still be used for finding initial mappings, but the dynamic nature of complex systems requires the mapping to be re-done during runtime in order to adapt to the new situation. To cope with such cases, new heuristics must be found so that their impact is minimal regarding (a) the information they need about the platform state, (b) the overhead caused by the re-mapping of the different parts of an application, and (c) the execution overhead of the heuristic itself.

Specific topic for PhD research:

Evolutionary algorithms have been currently used to statically optimise design and mapping of multiprocessor platforms. Timeliness and energy dissipation have been comprehensively addressed, but there is still scope for a further research on optimising other relevant aspects to multicore embedded systems such as reliability and security, or on coping with the issues raised by the latest semiconductor advances, such as variability and silicon aging.
As the computational power of such platforms keeps increasing, and so does the complexity and dynamism of applications they execute, it is natural to assume that evolutionary algorithms could also be employed during runtime to "evolve" optimised configurations and mappings as the dynamics of the system changes. Interesting research questions include the fine-tuning of the evolutionary approach (which could be done statically or also during runtime) and the decision on how much of the platform resources should be allocated to the evolutionary algorithm itself (how many cores, how much of the interconnect, how to bound its utilisation) so that its overhead will always be less than the optimisation benefits it provides.