面向大规模服务性能预测的在线学习方法
发布时间：2018-10-21 16:06
【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。
[Abstract]:In order to improve the quality of service, it is necessary to actively prevent service failure and service performance fluctuations, rather than trigger a handler when a service error occurs. Efficient prediction and analysis of the performance of large-scale services is an effective and feasible active prevention tool. However, the traditional service performance prediction model mostly adopts the full batch training model, which is difficult to meet the real-time requirements of large-scale service computing. On the basis of synthetically weighing the complete batch learning method and the stochastic gradient descent method, a large-scale service performance prediction model based on online learning is established, and a service performance prediction method based on small batch online learning is proposed. By reasonably setting the batch parameters of the prediction model, only the sample data of small batch size need to be trained in one iteration to improve the time efficiency of large-scale service performance prediction, and the convergence of online service prediction model is analyzed in detail. Experiments show that the proposed online learning algorithm can effectively solve the time-efficiency problem of large-scale service prediction algorithm.
【作者单位】： 南京航空航天大学计算机科学与技术学院;滁州学院地理信息科学系;上海第二工业大学计算机与信息学院;
【基金】：国家自然科学基金Nos.61672022,61272036 安徽省高等学校自然科学基金No.KJ2017A414~~
【分类号】：TP181
本文编号：2285633
[Abstract]:In order to improve the quality of service, it is necessary to actively prevent service failure and service performance fluctuations, rather than trigger a handler when a service error occurs. Efficient prediction and analysis of the performance of large-scale services is an effective and feasible active prevention tool. However, the traditional service performance prediction model mostly adopts the full batch training model, which is difficult to meet the real-time requirements of large-scale service computing. On the basis of synthetically weighing the complete batch learning method and the stochastic gradient descent method, a large-scale service performance prediction model based on online learning is established, and a service performance prediction method based on small batch online learning is proposed. By reasonably setting the batch parameters of the prediction model, only the sample data of small batch size need to be trained in one iteration to improve the time efficiency of large-scale service performance prediction, and the convergence of online service prediction model is analyzed in detail. Experiments show that the proposed online learning algorithm can effectively solve the time-efficiency problem of large-scale service prediction algorithm.
【作者单位】： 南京航空航天大学计算机科学与技术学院;滁州学院地理信息科学系;上海第二工业大学计算机与信息学院;
【基金】：国家自然科学基金Nos.61672022,61272036 安徽省高等学校自然科学基金No.KJ2017A414~~
【分类号】：TP181
【相似文献】
相关期刊论文 前3条
1 初元璋,祁鹏,张娅;宝钢IF钢大生产产品性能预测[J];北京科技大学学报;2001年01期
2 王丹民;李华德;郝春辉;;邯钢数据传输系统的建立及基于该系统的带钢性能预测[J];冶金自动化;2006年03期
3 刘振宇;周思杭;谭建荣;唐宏亮;;基于多准则修正的产品性能多参数关联分析与预测方法[J];机械工程学报;2013年15期
相关会议论文 前1条
1 陈超;李忠辉;熊昆鹏;;神经网络模型在沙钢热轧性能预测中的应用[A];全国冶金自动化信息网2014年会论文集[C];2014年
，本文编号：2285633
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2285633.html
