Ensembles of Models for Automated Diagnosis of System系统自动诊断模型的集合.pptVIP

下载本文档

2
0
约5.83千字
约 17页
2017-02-13 发布于上海
举报
版权申诉

Ensembles of Models for Automated Diagnosis of System系统自动诊断模型的集合.ppt

1、原创力文档（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。。
2、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
3、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。
4、该文档为VIP文档，如果想要下载，成为VIP会员后，下载免费。
5、成为VIP后，下载本文档将扣除1次下载权益。下载后，不支持退款、换文档。如有疑问请联系我们。
6、成为VIP后，您将拥有八大权益，权益包括：VIP文档下载权益、阅读免打扰、文档格式转换、高级专利检索、专属身份标志、高级客服、多端互通、版权登记。
7、VIP文档为合作方或网友上传，每下载1次，网站将根据用户上传文档的质量评分、类型等，对文档贡献者给予高额补贴、流量扶持。如果你也想贡献VIP文档。上传文档

Ensembles of Models for Automated Diagnosis of System Performance Problems Steve Zhang (steveyz@cs.stanford.edu) Armando Fox In collaboration with: Ira Cohen, Moises Goldszmidt, Julie Symons of HP Labs Motivation Complexity of deployed systems surpasses ability of humans to diagnose, forecast, and characterize behavior Scale and interconnectivity Levels of abstraction Lack of closed-form mathematical characterization Changes in input, applications, infrastructure Need for automated diagnosis tools Metric Attribution Service Level Objectives (SLO) Usually expressed in terms of high-level behaviors, e.g. desired average response time or throughput Which low level metrics are most correlated with a particular SLO violation Root cause diagnosis usually requires domain-specific knowledge Often does point to root cause Technique can be applied to more than just performance problems Background and Previous Work Modeling system behavior as a pattern classification problem Learn a function F that maps state of low level metrics (M) to SLO state (s+/s-) Tree Augmented Na?ve Bayes (TAN) models are used to represented F’s underlying probability distribution Models evaluated using Balanced Accuracy (BA) = 0.5(Prob(F(M) = s-|s-) + Prob(F(M) = s+| s+)) First step is feature selection Find subset of metrics that produces F with best BA Usually use heuristic search (greedy search) Avoids curse of dimensionality as well as overfitting Background (Cont.) TAN Models have key property of interpretability: (Metric Attribution) TAN models represent joint probability estimate P(s+/-,M) For each SLO violation, invert the joint probability estimate of each metric to obtain P(Mi|s+) and P(Mi|s-). Metric is attributed with an SLO violation if difference P(Mi|s+) - P(Mi|s-) 0. Output is a list of metrics that are implicated with each SLO violation. (out of the subset of the metrics chosen during the feature selection process) Not straightforward with many popular classifiers like Su