Appropriate allocation of workloads on performance asymmetric multicore architectures via deep learning algorithms

Abstract Asymmetric multicore processors (AMP) have become popular in both high-end and low-end computing systems due to its flexibility and high performance. A performance asymmetric multicore architecture (P-AMP) is the subcategory of AMP, which integrates the different micro-architecture cores in the same chip. Due to the heterogeneity nature of cores and applications, recognizing an optimal hardware configuration in terms of core, voltage-frequency pair for each application is still an NP-hard problem. Optimization of energy-delay product (EDP) is an additional challenging task in such architectures. To address these challenges, we developed a novel core prediction model called lightweight-deep neural network (LW-DNN) for asymmetric multicore processors. The proposed LW-DNN includes three phases, feature selection, feature optimization, and core prediction module. In the first and second phases, workload characteristics are extracted and optimized using the pre-processing algorithm and in the third phase, it predicts the appropriate cores for each workload at runtime to enhance the energy-efficiency and performance. We modeled a deep learning neural network using scikit-learn python library and evaluated in ODROID XU3 ARM big-Little performance asymmetric multicore platform. The embedded benchmarks we considered are MiBench, IoMT, Core-Mark workloads. The proposed LW-DNN prediction module compared with other traditional algorithms in terms of accuracy, execution time, energy consumption, and energy-delay product. The experimental results illustrate that accuracy achieved up to 97% in core prediction, and the average improvement in minimization of energy consumption is 33%, 35% in energy-delay product, 33% minimized in execution time correspondingly.