SHF: Small: Virtualizing Coordinated Resource Management of Flows on Handhelds with VIADUCT

手机、平板电脑和可穿戴设备等手持设备在无处不在的世界中得到迅速采用。相应地，用户在这些设备上部署的要求非常高的交互式应用程序如媒体流、交互式游戏、视频会议和社交网络也在激增。尽管这些设备中的单个组件的能效得到了极大的提高，但硬件组件的零碎优化不足以满足这些应用的苛刻需求，这些应用以非常复杂的方式同时使用多个CPU内核、加速器、存储系统、系统外围设备和传感器。 系统资源之间缺乏协调会导致运行此类应用时资源利用率低和系统级能效差。该项目没有投入更多的硬件来解决这个问题，而是提出了VIADUCT，它分配和协调系统级资源，以更节能的方式提高这些平台的并行性和资源利用率。认识到许多这些应用程序定期处理数据帧，这些数据帧需要在软实时范围内以流水线方式流过几个硬件组件，VIADUCT在应用程序中为每个此类帧流创建一个虚拟通道。该通道保存所有硬件组件之间分配的资源，并通过运行时协调来确保无缝的帧移动，而不会出现组件之间传统接口效率低下的情况。细粒度缓存和资源管理利用高级语义知识，以最节能的方式为每个流提供服务质量（QoS）。具体的研究贡献是在硬件接口，机制和政策链的硬件组件创建一个虚拟通道的流，架构的存储系统作为一个有效的管道，以维持这种流，并设计系统，以支持多个并发流。该项目旨在为这些平台提供一个全面的评估测试平台，以扩大研究界，并旨在让代表性不足的群体参与研究，并加强宾夕法尼亚州立大学？相关主题中的scrubbing。

Handheld devices such as cell phones, tablets and wearables, are gainingrapid adoption in a ubiquitous world. Correspondingly, there is anexplosion of very demanding and interactive applications such as mediastreaming, interactive games, video conferencing and social networking,that users are deploying on such devices. Despite tremendousimprovements in the energy efficiencies of individual components foundin these devices, piece-meal optimizations of hardware components areinadequate to accommodate the demanding needs of these applications thatemploy multiple CPU cores, accelerators, memory systems, systemperipherals and sensors, concurrently in very sophisticated ways. Lackof coordination among system resources leads to low resource utilizationand poor system-level energy efficiencies when running suchapplications. Rather than throw more hardware at this problem, thisproject proposes VIADUCT, which allocates and orchestrates system-levelresources to boost parallelism and resource utilization for theseplatforms in a more energy-efficient fashion.Recognizing that many of these applications periodically process framesof data that need to flow in a pipelined manner through several hardwarecomponents within soft real-time bounds, VIADUCT creates a virtualchannel per flow of such frames in an application. The channel holdsresources allocated across all the hardware components, with runtimecoordination to ensure seamless frame movements without the conventionalinterface inefficiencies across the components. Fine-grain schedulingand resource management leverages knowledge of high level semantics toprovide Quality-of-Service (QoS) for each flow in the mostenergy-efficient fashion. The specific research contributions are inhardware interfaces, mechanisms and policies to chain the hardwarecomponents for creating a virtual channel for a flow, architecting thememory system as an efficient conduit to sustain this flow, anddesigning the system to sustain several concurrent flows. The projectaims to provide a holistic evaluation testbed for such platforms amongstthe broader research community, and also aims to involveunder-represented groups in the research and enhance Penn State?scurriculum in related topics.