CRII: CSR: Online Performance Modeling of Opaque Cloud Applications

许多在线服务现在都托管在云上。然而，对于对性能敏感的用户来说，云的采用仍然非常有限。云用户面临的主要挑战之一是缺乏对云资源分配与应用程序性能之间关系的理解。虽然云服务提供商（CSP）为用户提供了轻松访问云资源以满足其计算需求的机会，但它们并不对用户部署的性能提供任何保证，也不提供用户应如何设置其资源分配的任何指导方针。这是因为用户部署是不透明的：CSP无法控制或访问用户的工作负载或应用程序。更糟糕的是，用户的云实例的有效容量可能会由于其他用户的干扰而动态变化。因此，云部署受到性能问题的困扰。这项研究的目标是开发新的性能模型，以帮助用户和CSP了解云应用程序的动态资源需求，而不需要任何广泛的基准测试或仪器。研究团队正在构建新颖的特定于工作负载的性能模型，以捕捉云资源分配和应用程序性能之间的关系。该团队还在开发基于控制理论和机器学习的新型在线工具，以动态推断（可能变化的）不可观察的云参数，从而允许在线调整性能模型。由此产生的模型将使用户和CSP能够准确地分配云资源，以实现所需的应用程序性能。基于云的服务正变得越来越受欢迎。该项目有助于提高云的采用率，并促进更有效地使用云。该研究提供了减少云资源浪费的工具，从而帮助云用户减少支出，并降低CSP数据中心的功耗。

Many online services are now hosted on the cloud. However, cloud adoption is still very limited when it comes to performance sensitive users. One of the primary challenges for cloud users is the lack of understanding of how cloud resource allocation relates to application performance. While Cloud Service Providers (CSPs) offer users easy access to cloud resources for their computing needs, they do not provide any guarantees on the performance of a user's deployment or any guidelines on how users should set their resource allocations. This is because user deployments are opaque: CSPs cannot control or access a user's workload or application. To make matters worse, the effective capacity of a user's cloud instance can change dynamically due to interference from other users. As a result, cloud deployments are plagued with performance issues. The goal of this research is to develop novel performance models to help users and CSPs understand the dynamic resource requirements of cloud applications without requiring any extensive benchmarking or instrumentation. The research team is constructing novel workload-specific performance models that capture the relationship between cloud resource allocation and application performance. The team is also developing novel online tools based on control theory and machine learning to dynamically infer the (possibly changing) unobservable cloud parameters, thus allowing the performance models to be tuned online. The resulting models will enable users and CSPs to accurately allocate cloud resources to achieve the desired application performance. Cloud-based services are becoming increasingly popular. This project helps improve cloud adoption and promotes more efficient use of the cloud. The research provides tools to reduce wastage of cloud resources, thus helping cloud users reduce their expenditure and also lowering the power consumption in CSP data centers.