Dynamic Task Allocation Strategies in MPSoC for Soft Real-time Applications

This work evaluates task allocation strategies based on bin-packing algorithms in the context of multiprocessor systems-on-chip (MPSoCs) with task migration capabilities, running soft real-time applications. The task migration model assumes that the whole code and data of the tasks are transferred from an origin node to the chosen destination node. We combine two types of algorithms to obtain better allocation results. Experimental results show that there is a trade-off between deadline misses and system energy consumption when applying bin-packing and linear clustering algorithms. In order to save energy, our system turns off idle processors and applies Dynamic Voltage Scaling to processors with slack. Depending on the algorithm selection and on the application, it is possible to obtain a reduction on deadline misses from 30 % to 100 % and energy consumption savings from 60 % to 80 %.