Deep neural network learning for power limited heterogeneous system with workload classification

Heterogeneous systems providing diverse computational capabilities have unlocked a new pathway in multicore processors. The versatility in applications and their ever-increasing performance demands have brought a paradigm shift to the heterogeneous systems. We use a deep neural network (DNN) based model to maximize performance under power constraints in heterogeneous systems. The dynamic power management technique is implemented in three stages. In the first stage, the core statistics and workload characteristics are collected for the DNN training at the later stage. This step dynamically estimates workload change for the current epoch using dynamic voltage frequency scaling (DVFS) based heuristic algorithm. In the second stage, DNN is trained through collected data points. The third stage uses the trained DNN model to identify a suitable voltage-frequency values to maximize performance under power capping. The power manager controls power-consumption at both per-core and per-chip levels. Our DNN prediction model is trained to address both core types (Large and Small) thus, improving the accuracy of the model. Simulations indicate that the proposed approach can achieve an overall 10.63% reduction in power-consumption and 5.5–6.8% improvement in power-savings when compared with the existing approaches. Besides, the proposed DNN model-based approach is able to maintain power capping with 95.81% accuracy with performance degradation of only 5.38% for a quad-core architecture.