SHF: Small: Energy Saving in Heterogeneous Data Centers

许多关键的在线服务正在转向云基础设施，以满足可扩展性需求。为了维持云计算的增长，有必要扩大数据中心的计算能力，提高能效。现代数据中心越来越多地将图形处理单元(GPU)等加速器与传统CPU集成在一起，以提供前所未有的并行性和数量级的计算吞吐量提升。为了在云规模上管理软件工作负载和硬件资源，数据中心正日益虚拟化，以提供轻松的软件和硬件管理。然而，现有的能效技术并不能很好地适用于虚拟化技术和异类硬件。该项目为新兴的虚拟异类数据中心实现高能效计算提供了基本见解和解决方案。该项目对国家利益的许多工作负载背后的计算引擎具有广泛的好处，例如天气预报和机器学习。这项研究的结果被整合到现有的本科生和研究生课程中。研究包括三个主要方面：1)开发异类基准来评估服务器能量分布和度量来量化异类服务器能量比例，以便调查虚拟化和异类体系结构的含义；2)开发用于容器实时迁移的软件支持和性能感知迁移策略以最小化实时迁移开销；3)开发管理技术以最大化异类CPU和多加速器系统的能量比例，例如动态负载平衡和自动运行时任务迁移。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Many critical online services are turning to cloud infrastructure to meet scalability demands. In order to sustain cloud computing growth, it is necessary to scale the computational capacity, and improve the energy efficiency, of data centers. Modern data centers increasingly integrate accelerators, such as Graphical Processing Units (GPUs), with traditional CPUs to provide unprecedented parallelism and order-of-magnitude improvement to computational throughput. In order to manage software workloads and hardware resources at cloud-scale, data centers are increasingly being virtualized to provide ease of software and hardware management. However, existing energy efficiency techniques are not well-suited for virtualization technology and heterogeneous hardware. This project provides fundamental insights and solutions towards achieving energy-efficient computing for emerging virtualized heterogeneous data centers. This project has wide-reaching benefits for the computational engines behind many workloads of national interest, such as weather forecasting and machine learning. Results of this research are being integrated into the existing undergraduate and graduate courses.The research consists of three main thrusts: 1) development of heterogeneous benchmarks to evaluate server energy profile and metrics to quantify heterogeneous server energy proportionality in order to investigate the implications of virtualization and heterogeneous architectures, 2) development of software support for container live migration and performance-aware migration strategies to minimize live migration overhead, 3) development of management techniques to maximize energy proportionality of heterogeneous CPU and multi-accelerator systems, such as dynamic load balancing and automatic runtime task migration.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.