一句话：提出 ACTGAN，用 GAN 学习优质配置的隐藏结构并生成更优配置，实现软件系统自动调优。
复杂软件系统通常提供大量参数供用户配置，但配置调优需要深入理解系统，远超普通用户能力。现有方法多关注学习性能估计模型，但在调优时间约束下样本稀少，模型精度受限。本文假设优质配置共享某些隐藏结构，提出 ACTGAN：不再仅提升对给定配置的性能预测，而是学习优质配置的隐藏结构并据此生成潜在更优配置。作者在 8 个软件系统的 17 种工作负载上评估，结果显示 ACTGAN 平均优于默认配置 76.22%，优于六种先进配置调优算法 6.58%–64.56%；生成的配置常优于训练样本且符合领域知识，验证了假设。
Complex software systems often provide a large number of parameters so that users can configure them for their specific application scenarios. However, configuration tuning requires a deep understanding of the software system, far beyond the abilities of typical system users. To address this issue, many existing approaches focus on exploring and learning good performance estimation models. The accuracy of such models often suffers when the number of available samples is small, a thorny challenge under a given tuning-time constraint. By contrast, we hypothesize that good configurations often share certain hidden structures. Therefore, instead of trying to improve the performance estimation of a given configuration, we focus on capturing the hidden structures of good configurations and utilizing such learned structure to generate potentially better configurations. We propose ACTGAN to achieve this goal. We have implemented and evaluated ACTGAN using 17 workloads with eight different software systems. Experimental results show that ACTGAN outperforms default configurations by 76.22% on average, and six state-of-the-art configuration tuning algorithms by 6.58%-64.56%. Furthermore, the ACTGAN-generated configurations are often better than those used in training and show certain features consisting with domain knowledge, both of which supports our hypothesis.
DOI 原文 ·
@inproceedings{paperbot5314,
title = {ACTGAN: Automatic Configuration Tuning for Software Systems with Generative Adversarial Networks},
author = {Liang Bao and Xin Liu and Fangzheng Wang and Baoyin Fang},
booktitle = {ASE},
year = {2019},
doi = {10.1109/ase.2019.00051}
}