CSR--PDOS: Autonomic Systems: Integrating Machine Learning with Computer Systems

计算机系统正在成为自身成功的牺牲品。与测试实验室相比，部署软件的用户、软件环境和使用模式要多得多，这导致了两个基本问题。首先，软件复杂性占系统成本的很大一部分——管理、配置、更新、性能调优和支持系统的成本超过了硬件的成本。其次，软件复杂性成为关键系统属性（如可靠性、性能和可用性）的一级限制。这项工作的目标是构建集成机器学习的原型软件系统，以简化操作，避免性能问题，并有效利用计算机资源。对于像web服务这样的分布式系统，关键目标是避免性能缺陷和有效利用计算资源。软件支持解决了配置软件和在出现问题时修复软件或环境的问题。自主网络诊断要求系统识别和关联事件，以达到对网络状态的诊断。这项工作将解决当相互不信任的各方共享机器学习模型时出现的信息泄露的安全问题，因为它们可能用于软件支持（软件供应商和用户）和网络诊断（竞争的网络提供商）。

Computer systems are becoming victims of their own success. Deployedsoftware has orders of magnitude more users, software environments,and usage patterns than it has in testing labs, which leads to twofundamental problems. First, software complexity accounts for a largefraction of system cost---the cost of administering, configuring,updating, performance tuning, and supporting systems is surpassing thecost of the hardware. Second, software complexity becomes afirst-order limit on key system properties like reliability,performance, and usability. The goals of this work are to build prototype software systems thatintegrate machine learning to simplify operation, avoid performanceproblems, and make efficient use of computer resources. Fordistributed systems like web services the key goals are avoidingperformance pitfalls and making efficient use of computing resources.Software support addresses the problem of configuring software andfixing it or the environment when something goes wrong. Autonomicnetwork diagnosis requires the system to identify and correlate eventsto reach a diagnosis for the state of the network. The work will addressthe security problems of information leakage that arise when mutuallydistrustful parties share machine learning models, as they might forsoftware support (software vendor and user) and network diagnosis(competing network providers).