Real-time optimization of energy consumption and load balancing in server clusters based on model decomposition and a new differential evolution variant
Zhi Xiong, Ziyu Wang, Jianlong Xu, Fei Wang · Swarm and Evolutionary Computation · 2025
Adjusting server deployment in a server cluster in real time with a changing workload to balance energy saving and load balancing is an urgent problem. Addressing the limitations of existing research in decision variable reduction and constraint handling, this work proposes an online real-time optimization strategy for energy consumption and load balancing of server clusters based on model decomposition and a new differential evolution (DE) variant. The optimization content includes the on/off state, CPU frequency, and workload allocation of each server. The decision variables are reasonably defined to derive the cluster energy consumption and load balancing models, and the cluster optimization is described as a bi-objective optimization model. Then, based on the model characteristics, the model is decomposed into two layers of optimization to reduce the difficulty of solving the problem, where the outer layer is a bi-objective nonlinear optimization problem and the inner layer is a single-objective mixed-integer linear programming problem. Finally, the inner-layer optimization is solved using the Gurobi optimizer, and the outer-layer optimization is solved using the non-dominated sorting genetic algorithm II and DE algorithm. To address the constraints existing in the outer-layer optimization, a new DE variant, DE/rand/1/while-if-either-or, is proposed. This variant can increase the feasible probability of mutant individuals and reduce the interference with the evolution mechanism, improving the population quality. Tests in various scenarios verify the real-time optimization capability of the proposed strategy and demonstrate the feasibility and effectiveness of the model decomposition and the new DE variant.