Investigation of LSTM based prediction for dynamic energy management in chip multiprocessors

In this paper, we investigate the effectiveness of using long short-term memory (LSTM) instead of Kalman filtering to do prediction for the purpose of constructing dynamic energy management (DEM) algorithms in chip multi-processors (CMPs). Either of the two prediction methods is employed to estimate the workload in the next control period for each of the processor cores. These estimates are then used to select voltage-frequency (VF) pairs for each core of the CMP during the next control period as part of a dynamic voltage and frequency scaling (DVFS) technique. The objective of the DVFS technique is to reduce energy consumption under performance constraints that are set by the user. We conduct our investigation using a custom Sniper system simulation framework. Simulation results for 16 and 64 core network-on-chip based CMP architectures and using several benchmarks demonstrate that the LSTM is slightly better than Kalman filtering.