Abstract: To improve the quality of cloud services, service performance violations need to be proactively prevented instead?of?recovery triggered by the occurrence?of?failures. The performance predicting model is a promising tool for evaluating the status of Web services in cloud computing. However, traditional batch machine learning techniques could not satisfy the requirement of real-time predicting in large-scale service-oriented applications. To deal with the challenges, this paper proposes a mini-batch online learning approach to predict the performance of large-scale services. Through properly setting the batch parameters, the proposed approach uses a small fixed size of samples to train the prediction model in each iteration. This strategy efficiently reduces the computations per iteration. Furthermore, detailed theoretical analysis is conducted for online learning models and their algorithm convergence. The experimental results indicate that the proposed approach is feasible and effective.

Key words: large-scale service computing, online learning, mini-batch online learning, stochastic gradient descent

摘要： 为提高服务运行质量，需要主动预防服务失效和服务性能波动，而不是在服务发生错误时触发处理程序。高效地预测与分析大规模服务的性能是有效可行的主动预防工具。然而传统的服务性能预测模型多采用完全批量训练模式，难以满足大规模服务计算的实时性要求。在综合权衡完全批量学习法和随机梯度下降法的基础上，建立了基于在线学习的大规模服务性能预测模型，提出了一种基于小批量在线学习的服务性能预测方法，通过合理地设置预测模型的批量参数，一次迭代仅需训练批量规模较小的样本数据，从而改善大规模服务性能预测的时间效率；详细分析了在线服务预测模型的收敛性。实验表明，提出的在线学习算法有效地解决了大规模服务预测算法的时效性问题。

关键词: 大规模服务计算, 在线学习, 小批量在线学习, 随机梯度下降法