ITR/SY: QoS-Aware Execution Environments for Multi-tier Network Services

现代的组件化Internet服务通过共享公共服务组件允许服务器合并。这种共享可以潜在地降低安装成本、空间和能量消耗。 然而，迄今为止，服务很少在服务器应用程序之间共享，以避免它们之间可能的干扰。 服务提供者经常运行多个共享服务的实例，每个前端服务一个。到目前为止，从多层服务的部署中获得的唯一好处是可以使用现有的服务组件快速组合新服务。由于这些服务组件可以在不同的机器上执行，每个最佳配置为它的服务，多层服务往往比monolithic people.拟议的研究旨在解决当前操作系统在执行多层服务的缺点。我们将重点介绍跨多层服务器群的层关联和跟踪服务活动的操作系统扩展。我们将开发活动在线分类机制（例如，溢价与~基本服务类），以及允许系统管理员隔离（性能方面）不同服务类的机制，即使不同服务类的活动可能重叠并共享后端服务。用于管理多层服务的相同操作系统机制可用于分析不同层之间的交互以及不同服务类别之间的干扰。我们将表明，服务交互分析和服务警务的组合，最大限度地减少服务/应用程序之间的性能入侵。我们还将研究如何在这种环境中改进服务之间的隔离。我们将探讨服务实现如何利用改进的操作系统基础架构来实现多层

Modern, componentized Internet services allow server consolidation by sharing common service components. Such sharing can potentially reduce the installation cost, space, and energy consumption. However, to date, services are seldom shared across server applications to avoidpossible interference between them. Service providers often runmultiple instances of shared services, one for each frontendservice. The only benefit gained thus far, from the deployment ofmulti-tier services is that new services can be composed quickly usingexisting service components. Since these servicecomponents may execute on different machines, each optimally configuredfor its services, multi-tiered services tend to be morescalable than monolithic counterparts.The proposed research aims to address the shortcoming of currentOSs in executing multi-tiered services. We will focus on OS extensions that correlate and trace service activities across the tiers of a multi-tiered server farm. We will develop mechanisms for onlineclassification of activities (e.g., premium vs.~basic service classes), as well as mechanisms that allow system administrators to insulate (performance-wise) different service classes even when the activities of different service classes may overlap and share backend services. The same OS mechanisms that are used to manage multi-tiered services can be used to analyze the interactions between different tiers and the interferences between different service classes. We will show that a combination of service interaction analysis and service policing minimizes performance invasiveness between services/applications. We will also investigate how to improve insulation between services in this environment. We will explore how service implementations may take advantage of an improved OS infrastructure for multi-tiered