Traditional stress testing methods are difficult to simulate the complexity and dynamics in real business scenarios, resulting in distorted test results and low efficiency. In order to solve the above problems, this paper proposes an automated framework for stress testing of AI servers facing high concurrency scenarios. The framework adopts the design concept of hierarchical decoupling and intelligent decision-making, and consists of four modules: intelligent load generation layer, system resources and performance monitoring layer, dynamic tuning and control center, root cause analysis and report generation layer. Among them, the intelligent load generation layer supports mixed simulation of multi-modal AI loads, the dynamic tuning and control center realizes dynamic optimization of test parameters based on reinforcement learning (RL) algorithm, and the root cause analysis and report generation layer automatically locates performance bottlenecks and generates reports by unsupervised learning and time series correlation analysis. The experimental results show that the framework can effectively find the potential bottlenecks of the system, improve the test efficiency, and shorten the fault diagnosis cycle, which provides strong support for the performance optimization of AI server.
Xingcheng Ren (Fri,) studied this question.