CSR---PDOS: A Holistic Framework for Power and Performance Control in Data Centers

近年来，托管数千台高密度刀片服务器并提供外包商业IT服务的企业数据中心快速增长。有效运营现代数据中心有两个关键挑战。首先，不同的客户必须满足他们所需的服务级别协议(SLA)，例如响应时间和吞吐量。其次，为了降低运营成本，必须控制功耗，避免因电力容量过载或系统过热而导致的故障。本项目旨在开发一种全面的管理框架，同时控制功耗和所需的应用层SLA，并为两者提供理论保障。此跨层框架采用受限的多输入多输出控制模型，可在群集级别管理多台刀片服务器。与依赖启发式管理方案的传统解决方案形成鲜明对比的是，本研究采用了一种严格的设计方法，该方法以控制理论为基础，系统地制定控制策略，并从分析上保证控制精度和系统稳定性。该项目还为在数据中心的大量服务器上运行的大型应用服务开发分散控制算法。此外，还将解决系统可控性保证和热管理等相关挑战。

Recent years have seen a rapid growth of enterprise data centers that host thousands of high-density blade servers and provide outsourced commercial IT services. There are two key challenges for effectively operating a modern data center. First, different customers have to be assured by meeting their required service-level agreements (SLAs) such as response time and throughput. Second, power consumption has to be controlled in order to reduce operation costs, and avoid failures caused by power capacity overload or system overheating due to increasing high-density.This project aims to develop a holistic management framework which controls power consumption and required application-level SLAs simultaneously with theoretical guarantees on both. This cross-layered framework features a constrained multi-input-multi-output control model to manage multiple blade servers at the cluster level. In sharp contrast to traditional solutions that rely on heuristic-based management schemes, this research adopts a rigorous design methodology that is based on a control-theoretic foundation for systematically developing control strategies, with analytic assurance of control accuracy and system stability. This project also develops decentralized control algorithms for large-scale application services running on a large number of servers in a data center. In addition, related challenges such as system controllability guarantee and thermal management will be addressed.