SHF: Small: Optimizing Consolidation Efficiency of Emerging Virtualized Cloud Applications on Contemporary Server Architecture

优化虚拟机（VM）整合性能一直是云提供商社区面临的最关键任务之一。然而，当扩展虚拟平台以处理云应用程序的激增时，由于性能下降的限制，特别是在VM整合密度持续增长的环境中，会出现挑战。更糟糕的是，整合的工作负载正在从传统的面向计算的单任务应用程序转移到大规模和复杂的工作负载。它们产生许多不同的交互和通信模式，其中一些应用程序具有大量不规则的内存占用和高内存消耗，这对优化虚拟机整合的效率提出了重大挑战。目前，出现了一些新的技术，为当代服务器架构。服务器制造商正在用非统一内存访问（NUMA）机器取代传统的统一内存访问（UMA）机器，因为后者具有更高的内存带宽和更好的系统可扩展性。另一方面，面向吞吐量的图形处理单元（GPU）越来越多地部署到云数据中心服务器以满足计算需求。现代虚拟机管理程序开始虚拟化GPU资源，并为托管的虚拟机提供高效可靠的性能。为了抓住机遇并应对相关挑战，该研究项目将开发技术，以提高新兴虚拟化云应用程序的整合效率，同时继续推进底层服务器架构。项目目标包括：（1）NUMA架构的系统范围整合性能分析和优化;（2）图形即服务（GaaS）工作负载整合开销表征和最小化;以及（3）NUMA和GPU服务器配置方面的集体工作负载整合优化。该项目以协同方式集成了新兴的服务器架构功能和虚拟机管理程序资源管理，将为云和大数据计算的新型高效横向扩展计算平台打开大门。它将有助于使计算系统保持其历史性的扩展，从而使许多现实生活中的应用程序受益。该项目还将通过让代表性不足的群体参与进来，以及为教育和培训传播研究基础设施，为社会做出贡献。

Optimizing Virtual Machine (VM) consolidation performance has been one of the most critical tasks faced by the cloud provider community. Nevertheless, when scaling virtual platforms to handle the proliferation of cloud applications, challenges arise due to constraints on performance degradation, especially in those environments where VM consolidation density continues to grow. Worse, the consolidated workloads are shifting from conventional single-task computation-oriented applications to large-scale and complex workloads. They generate many diverse interactions and communication patterns and some of these applications have a large irregular memory footprint and high memory consumption, which presents a significant challenge to optimizing the efficiency of virtual machine consolidation. Currently, there emerge some new techniques for contemporary server architecture. Server manufacturers are replacing traditional Uniform Memory Access (UMA) machines with Non-Uniform Memory Access (NUMA) ones due to the latter's higher memory bandwidth and better system scalability. On the other hand, the throughput-oriented graphics processing units (GPUs) are being increasingly deployed to cloud data center servers to meet computation demands. Modern hypervisors begin to virtualize GPU resources and deliver efficient and reliable performance to hosted virtual machines. To embrace the opportunities and address the associated challenges, this research project will develop techniques to improve the consolidation efficiency of emerging virtualized cloud applications as the advancement of the underlying server architecture continues. The project objectives include: (1) System-wide consolidation performance profiling and optimization for NUMA architecture; (2) Graphic-as-a-Service (GaaS) workload consolidation overhead characterization and minimization; and (3) Collective workload consolidation optimizations in terms of both NUMA and GPU server configurations. This project, which synergistically integrates emerging server architecture features and virtual machine hypervisor resource management, will open the door for a new class of efficient scale-out computing platforms for cloud and big data computing. It will contribute to enabling computing systems to stay on track with its historic scaling and hence benefit numerous real-life applications. This project will also contribute to society through engaging under-represented groups, and research infrastructure dissemination for education and training.