Online DVFS using Deep Learning: Sequence to Sequence LSTM Networks with Attention

Real Time embedded systems are highly complex due to interactions and interdependencies between various hardware/software units and policies of the processors with applications running on it. To deal with fluctuating workloads and subsequent tasks, smart adaptability of supply clock and voltage is required in order to optimize power without compromising on the performance. This is done using Dynamic Voltage and Frequency Scaling (DVFS) technique. An improved version of DVFS is proposed in this paper which treats it as a recurrent problem with an aim to capture the intricate dependencies amongst various factors influencing the operation. The authors have employed application independent- Radial Basis Neural Network to generate series of predicted frequencies for current workload of the processor, followed by seq2seq-LSTM based encoder decoder model using Attention to decide if the frequency generated by the Artificial Neural Network (ANN) model is optimum from power conservation point of view. The proposed model predicts the workload and then compares the predicted frequency to the critical value or deadline of the current task. The experiments were conducted on a single core processor on which a benchmark application was run, and promising prediction accuracy rates were obtained without incurring degradation of critical performance parameters.