一种基于深度学习日志分析模型的压力测试算法

A Stress Testing Algorithm Based on Deep Learning Log Analysis Model

摘要: 软件新系统上线前，常常需要对生产运行的系统开展压力测试，以规避系统因突发巨量客户端访问或生产过程中多用户集中操作，而出现报错及系统崩溃等严重问题。本文提出了一种基于深度学习日志分析模型的压力测试算法，该算法基于现有生产日志数据创建双向长短期记忆循环神经网络(Bi-LSTM)的分析模型，搭建并预测基于生产日志规律的闭环测试场景，模拟预测生产环境的理论测试场景，作为传统基于规则压力测试方法的补充。实验结果表明，使用本文所提双向LSTM分析模型可以有效地模拟生产日志的规律，使得压力测试更贴近真实场景，能有效观察系统运行状况，多维度了解系统性能的极限和缺陷。

Abstract: Before launching a new software system, stress testing on the production-environment system is necessary. It could avoid serious problems such as system errors and system crashes caused by sudden massive client access or concurrent operations by multiple users during the production process. This paper proposes a deep learning log analysis model-based stress testing algorithm. The algorithm first creates a bi-directional long and short-term memory recurrent neural network (Bi-LSTM) analysis model based on existing production log data, then builds and predicts a closed-loop test scenario based on production log patterns, and finally simulates a theoretical test scenario of the production environment. This method can be served as a supplement to the traditional rule-based stress testing method. The experimental results show that using the proposed bi-directional LSTM analysis model in this paper can effectively simulate the pattern of production logs, which makes the stress testing closer to the real scenario. In this case, developers can effectively observe the system operation condition and understand the limits and defects of the system performance in multiple dimensions.