面向IaaS云性能保证的资源弹性配置及其性价比优化研究

With the ability of offering tenants elastic computing resources on demand, while charging them by a simple pay-as-you-go pricing model, Infrastructure-as-a-Service (IaaS) cloud computing has recently gained much popularity and extensive development all over the world. However, the performance variation of virtual machines and the resource-centric service model can adversely affect the service capability of the IaaS cloud. Existing studies fail to effectively deliver predictable performance to tenant applications, due to the inaccuracy of application performance prediction model and the heavy overhead of resource provisioning methods, etc. To tackle this issue, we comprehensively evaluate the performance variation of various representative cloud applications across a number of popular IaaS cloud platforms. Then, we develop an online performance prediction model of cloud applications using the multi-dimensional resource utilization of virtual machines. In particular, our performance model takes into account the factor of performance variation caused by the hardware heterogeneity and performance interference of virtual machines. Based on such a performance model and the pricing of cloud resources, we further devise a cost model of resource provisioning with the aim of providing predictable performance to cloud applications. The deep analysis of such a cost model enables the design of a cost-effective elastic resource provisioning algorithm, which guarantees the performance of cloud applications and cuts down the monetary cost of resource provisioning. Accordingly, the results of our work can be used to jointly optimize the predictable performance and resource provisioning cost of tenant applications, and improve the service capability of the IaaS cloud, thereby achieving a win-win situation between cloud providers and tenants.

基础设施即服务（IaaS）云计算可为用户提供按需可扩展及现收现付式的计算资源，近年来在国内外已得到较为广泛的关注与发展。然而，虚拟机的性能波动问题和按资源租用的服务模式使得当前IaaS云计算的服务能力较为薄弱，已有研究由于应用性能预测不准确及资源配置方法开销大等问题，仍未能有效地为云应用提供性能保证。为此，本课题将全面评测分析主流IaaS云计算环境下多种典型云应用的性能波动影响，利用虚拟机多维度资源利用率构建云应用的在线性能预测模型，并综合加入考虑虚拟机硬件异构与性能干扰所造成的性能波动因素。通过该性能模型与云计算资源的价格体系，构建分析基于云应用性能保证的资源配置成本模型，进而设计一种优化虚拟机资源性价比的弹性配置算法，以保证云应用的性能，并降低资源配置成本开销。本课题的研究成果可用于实现云应用性能保证与资源配置成本开销的同步优化目标，提升IaaS云计算的用户服务能力，达到云服务商与用户的双赢。

在基础设施即服务（IaaS）云计算平台中部署运行用户应用负载已成为一大趋势。然而，云虚拟机的性能波动问题和按资源租用的服务模式使得当前IaaS云计算的服务能力较为薄弱，已有研究由于应用性能预测不准确及资源配置方法开销大等问题，仍未能有效地为云应用提供性能保证。为此，本课题将首先全面评测分析主流IaaS云计算环境下多种典型云应用的性能波动影响。评测结果表明，在云虚拟机中运行应用的性能波动可达3.9%-92.1%；其次，利用虚拟机多维度资源利用率构建云应用的在线性能预测模型，并综合加入考虑虚拟机硬件异构、性能干扰、以及由瞬时实例撤回所造成的性能波动影响。实验结果表明，性能预测模型可准确预测典型云应用（如MapReduce、Spark大数据分析应用）的运行性能，平均误差为6.1%-10.6%。最后，依据云应用性能模型和瞬时实例，进而设计一种优化虚拟机资源性价比的弹性配置算法，以保证云应用的性能并降低资源配置成本开销。实验结果表明，性价比优化资源配置算法可节约用户作业运行成本高达83.8%。本课题的研究成果可用于实现云应用性能保证与资源配置成本开销的同步优化目标。