Optimizing Large-Scale AI Server Testing via Adaptive Evolutionary Algorithms

Abstract

Traditional testing methods are inefficient in high-dimensional parameter space, dynamic load fluctuation and heterogeneous hardware architecture, and it is difficult to meet the testing requirements of large-scale clusters. In this paper, an Environment-Aware Adaptive Hybrid Algorithm is proposed, which can effectively solve the convergence stagnation problem of high-dimensional parameter space through dynamic mutation rate adjustment and hybrid coding strategy. EA-AHA adopts a master-slave island model, in which the master island is responsible for global exploration and the slave island is responsible for local exploitation, and maintains population diversity through individual migration. The algorithm uses chromosome representation coded by real numbers and integers, and uses adaptive mutation rate and crossover strategy to adapt to different test scenarios. The multi-objective fitness function comprehensively considers task execution time, system resource utilization rate and fault detection rate, so as to realize the synergistic improvement of test efficiency, resource utilization rate and fault detection rate. In addition, EA-AHA models the optimization problem as a dynamic optimization problem, and responds to dynamic changes such as background load fluctuation and hardware performance attenuation in the test environment through environmental state awareness and adaptive response mechanism. The closed-loop mechanism of simulation and physical verification further improves the optimization efficiency and practical effectiveness. The experimental results show that EA-AHA is superior to Bayesian optimization and NSGA-II algorithm in convergence speed, comprehensive performance and dynamic environment adaptability, which effectively solves the problem of large-scale AI server test optimization.