Fine-grained monitoring for self-aware embedded systems

Dynamic Thermal and Power Management methods highly depend on the quality of the monitoring, which needs to provide estimations of the system's state. This can be achieved with a set of performance counters that can be configured to track logical events at different levels. Although this problem has been addressed in the literature, recently developed highly reactive adaptation techniques require faster, more accurate and more robust estimations methods. A systematic approach (PESel) is proposed for the selection of the relevant performance events from the local, shared and system resources. We investigate an implementation of a neural network based estimation technique which provides better results compared to related works. Our approach is robust to external temperature variations and takes into account dynamic scaling of the operating frequency. It achieves 96% accuracy with a temporal resolution of 100 ms, with negligible performance/energy overheads (less than 1%).