Application Workload Characterization using BAT_LSTM Learning algorithm for Asymmetric Architectures

Nowadays, asymmetric multicore architectures become everywhere due to its energy efficiency, QoS, and high performance. Though workload characterization on these architectures become a challenging task due to its heterogeneous pipeline structure and execution process that affects the overall performance of the system. To resolve this issue, BAT_LSTM deep learning predictor has been designed and developed to predict appropriate resource for each workload at runtime. Deep learning algorithms are adopted in several applications such as computer vision, smart vehicles, and medical environment in order to classify and predict the unknown. In this work, BAT_LSTM neural network predictor has been designed and compared with random forest algorithms, decision tree, naive bayes and support vector machine for workload characterization. Cost functions of these algorithms are designed and developed in order to detect the optimal processor for each workload execution at runtime. Core mark workloads are initially executed on quad core multicore hardware to analyze the workload characteristics in terms of memory consumption, I/O, CPU usage, instructions type, cache miss ratios and so on. These characteristics are feed forwarded into machine a learning algorithm that identifies the best processor. Performance of proposed algorithms is evaluated using testing workloads in terms of processor prediction accuracy, execution time metrics. Average of 10% in energy consumption reduction and 96.8% in accuracy is achieved through proposed predictors.