Date of Award
Doctor of Philosophy (PhD)
Electrical and Computer Engineering
In this dissertation, we study dynamic reliability management (DRM) and dynamic energy management (DEM) techniques for network-on-chip (NoC) based chip multiprocessors (CMPs). In the first part, the proposed DRM algorithm takes both the computational and the communication components of the CMP into consideration and combines thread migration and dynamic voltage and frequency scaling (DVFS) as the two primary techniques to change the CMP operation. The goal is to increase the lifetime reliability of the overall system to the desired target with minimal performance degradation. The simulation results on a variety of benchmarks on 16 and 64 core NoC based CMP architectures demonstrate that lifetime reliability can be improved by 100% for an average performance penalty of 7.7% and 8.7% for the two CMP architectures. In the second part of this dissertation, we first propose novel algorithms that employ Kalman filtering and long short term memory (LSTM) for workload prediction. These predictions are then used as the basis on which voltage/frequency (V/F) pairs are selected for each core by an effective dynamic voltage and frequency scaling algorithm whose objective is to reduce energy consumption but without degrading performance beyond the user set threshold. Secondly, we investigate the use of deep neural network (DNN) models for energy optimization under performance constraints in CMPs. The proposed algorithm is implemented in three phases. The first phase collects the training data by employing Kalman filtering for workload prediction and an efficient heuristic algorithm based on DVFS. The second phase represents the training process of the DNN model and in the last phase, the DNN model is used to directly identify V/F pairs that can achieve lower energy consumption without performance degradation beyond the acceptable threshold set by the user. Simulation results on 16 and 64 core NoC based architectures demonstrate that the proposed approach can achieve up to 55% energy reduction for 10% performance degradation constraints. Simulation experiments compare the proposed algorithm against existing approaches based on reinforcement learning and Kalman filtering and show that the proposed DNN technique provides average improvements in energy-delay-product (EDP) of 6.3% and 6% for the 16 core architecture and of 7.4% and 5.5% for the 64 core architecture.