A Review on Workload Characteristics for Multi Core Embedded Architectures using Machine Learning and Deep Learning Techniques

Abstract

Optimization of energy consumption in system on-chip (SoC) become a challenging task in real-time. Detection of best core for each workload is an additional critical task. To overcome these challenges, we analyzed the different machine learning and deep learning algorithms for mapping each workload on the Quad-core platforms. In this paper, we adopted support vector machines (SVM), naïve baize, random forest and KNN for prediction of the best core for each workload. Initially, we observe the workload characterization in terms of memory, instruction cycles, branch data's and developed as a database. In second phase, we deployed the ML and DL algorithms with trained database to predict accurately the best core for each workload. In the third phase, prediction accuracy, energy consumption metrics are observed and compared with the traditional algorithms. The proposed model is executed on Rasp -pi Quad-Core hardware platform and ML algorithms are simulated on the python IDE. Simulation results illustrates the prediction accuracy is achieved up to 99% on LSTM prediction for IomT benchmark and 8.6% on the energy consumption metric when compared to other ML techniques.