资源虚拟化环境中面向I/O密集型负载的能效优化策略研究

With the introduction of cloud computing, resource virtualization technology provides several solutions for energy-efficiency optimization in datacenters. However, recent studies show that the energy-efficiency related metrics will be significantly reduced when the system is in presence of I/O-intensive workloads. The major reasons include: (1) When I/O-intensive and CPU-intensive workloads are coexisting, it is a challenging issue for virtual machine scheduler that both performance and fairness are considered at the same time; (2) Massive intermediate data result in higher I/O related energy consumption. To address the above issues, we firstly introduce the ‘Performance-Monitor-Counters Ratio Model’ to describe the power consumption of virtual machines, with aiming at increasing the accuracy of measuring virtual machine’s power consumption; then, we plan to implement an ‘I/O Compensating Mechanism’, which is aiming to optimize virtual machine scheduler in terms of energy-efficiency and execution performance; finally, we plan to develop a ‘Two-Phase Virtual Machine Deployment’ policy for improving the energy-efficiency when running large-scale I/O-intensive applications. The research results and related theories in this project are expected to be applied in virtualized cloud systems so as to reduce the energy-related costs in their high-performance datacenters.

随着云计算概念的推广，虚拟化技术为数据中心的能效优化提供了若干新的解决思路。但近期研究显示，在执行I/O密集型负载时，虚拟化平台的各类能效指标将显著降低，其主要原因在于：（1）I/O密集与CPU密集型任务共存时，虚拟机调度器难以兼顾效率和公平；（2）大量“中间数据”的I/O访问与管理显著增加了虚拟化平台的能耗开销。针对以上问题，本课题首先以研究虚拟机功耗模型为起点，提出采用“性能计数器比例模型”来描述虚拟机的功耗状态，以此提高虚拟机功耗度量的准确性；然后，通过高精度功耗模型来量化分析I/O密集型负载对资源虚拟层能效的负面影响，并研究如何通过“I/O补偿机制”来优化虚拟机调度器的能效表现；最后，研究如何通过两阶段虚拟机部署策略来降低“中间数据”所导致的I/O能耗开销。本课题的研究结论和相关理论将有望应用于采用虚拟化技术部署资源的云平台，以期降低其数据中心的能耗成本。

随着云计算系统的普及，资源虚拟化环境的能效优化已经成为各类数据中心必须面对的重要问题。能效优化不仅能够降低数据中心的运营成本，也能够提高系统的可靠性和稳定性。本课题主要针对虚拟化环境中I/O（数据）密集型负载，对此类负载在能效优化方面的难点和挑战提出相应的解决方案和思路。课题组首先利用PMC机制构建了一个可以实时测量各类虚拟机的功耗模型，并利用该功耗模型改进了现有虚拟化平台的本地调度算法，依此降低系统面对混合型负载时的能耗开销；随后，课题组针对数据密集型工作流调度、虚拟资源供给、虚拟内存分配、QoS优化、云存储等方面所涉及的能效优化问题提出了若干相应的解决方案和策略。这些研究方案和策略可以直接应用于现有的各类云数据中心，以此提高现有云数据中心的能效指标。项目研究成果包括：7篇已发表论文和6篇已录用论文，其中SCI检索1篇，EI检索5篇，培养青年教师4名。