Putting Hard Real-time Applications together with Dark Silicon

Sammanfattning: As the number of transistors per unit square on silicon chips continues to in- crease, the heat production of these chips also increases, making thermal problems more common. As the heat production is dependent on the clock frequency, chips today have a defined thermal safe speed at which the cooling system is sufficient to compensate for the produced heat. A DVFS enabled processor executing at speeds over the thermal safe speed for prolonged periods of time may be forced to power down a set of cores to prevent damage due to overheating. As any deadline violations are considered as a system failure, executing hard real-time workloads on such a platform is not trivial. This thesis presents an online scheme which enables the execution of workloads on such a system. The scheme uses a history aware bounding function for determining when the processor speed must be increased above the thermal safe speed, and as such has a constant and small memory and computation footprint. The scheme is analysed using predefined traces within the gem5 simulator. For this purpose a new simulated hardware device is implemented which allows the scheduler to read predefined trace files. With workloads that follow the model closely the online scheme performs on-par with the optimal offline implementation. For workloads whose arrival curve follow the model less closely, the presented scheme is still able to assign timing correct processor speeds.