CNS Core: Medium:Model-driven Resource Management for Avoiding Performance Pitfalls in Edge Computing

CNS 核心：中：模型驱动的资源管理，以避免边缘计算中的性能陷阱

• 批准号: 2211888
• 负责人: Prashant Shenoy
• 金额: $ 119.91万
• 依托单位: University of Massachusetts Amherst
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-10-01 至 2026-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2211888&HistoricalAwards=false
• 关键词: CNS; Core; Medium; Model; driven

A variety of new distributed applications are emerging that have strict low-latency requirements, including real-time machine learning-based inference, Internet-of-Things services, mobile augmented reality, and cloud gaming. To support these applications, cloud providers are building out large-scale distributed edge infrastructures that can provide computing and storage resources in close proximity to end users. Yet, despite the significant network latency advantage of edge servers, edge computing remains vulnerable to numerous performance pitfalls that can lead to reduced performance. This counter-intuitive behavior primarily occurs when edge resource constraints or workload bursts cause high queuing delays and response times that significantly increase latency. To address this problem, this project will develop rigorous analytical models of edge and cloud performance to gain a fundamental understanding of when and why edge performance problems occur in practice. The project will then apply these models to design novel, but practical, resource management policies that can enable edge computing to provide low latency for a wide range of real-world applications. These policies include i) edge elasticity and bursting that adaptively scale edge resources within and across edge and cloud sites under workload spikes; ii) performance isolation for edge accelerators that flexibly multiplexes accelerators across applications to increase their utilization; and iii) dynamic resource provisioning and allocation for serverless edge computing that determines the resources needed by serverless containers to satisfy tail latency requirements. Collectively, these models and policies will enable edge computing to fulfill its potential to support new classes of low-latency applications.The project has the potential for significant practical impact by enabling commercial cloud providers to offer low-latency edge cloud services, which is important in supporting numerous emerging applications with strict low-latency requirements. The project will conduct outreach by incorporating relevant research topics within summer programs for local middle and high school students. The project will also inject elements of edge computing and cloud computing into current graduate and advanced undergraduate classes at the PIs' institution. The project will emphasize recruiting a diverse group of undergraduate and graduate students through participation in REU programs and institutional diversity efforts. Finally, the software artifacts and datasets from the project will be made available to the research community as open source via the UMass Trace Repository.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.

各种新的分布式应用正在出现，它们具有严格的低延迟要求，包括基于实时机器学习的推理、物联网服务、移动的增强现实和云游戏。为了支持这些应用程序，云提供商正在构建大规模的分布式边缘基础设施，这些基础设施可以在终端用户附近提供计算和存储资源。然而，尽管边缘服务器具有显著的网络延迟优势，但边缘计算仍然容易受到许多性能缺陷的影响，这些缺陷可能导致性能下降。这种违反直觉的行为主要发生在边缘资源约束或工作负载突发导致高排队延迟和响应时间，从而显著增加延迟时。为了解决这个问题，该项目将开发边缘和云性能的严格分析模型，以从根本上了解边缘性能问题在实践中何时以及为什么会发生。然后，该项目将应用这些模型来设计新颖但实用的资源管理策略，使边缘计算能够为广泛的实际应用提供低延迟。这些策略包括i）边缘弹性和突发，其在工作负载峰值下自适应地扩展边缘和云站点内和跨边缘和云站点的边缘资源; ii）边缘加速器的性能隔离，其跨应用灵活地复用加速器以提高其利用率;以及iii）无服务器边缘计算的动态资源供应和分配，其确定无服务器容器满足尾部延迟要求所需的资源。总的来说，这些模型和策略将使边缘计算能够实现其支持新型低延迟应用的潜力。该项目通过使商业云提供商能够提供低延迟边缘云服务，对支持具有严格低延迟要求的众多新兴应用非常重要，具有重大的实际影响。该项目将通过将相关研究课题纳入当地初中和高中学生的暑期课程来进行推广。该项目还将把边缘计算和云计算的元素注入PI机构目前的研究生和高级本科生课程。该项目将强调通过参与REU计划和机构多样性努力招募多样化的本科生和研究生群体。最后，该项目的软件工件和数据集将通过麻省大学跟踪存储库作为开源提供给研究社区。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。

项目成果

Jointly Managing Electrical and Thermal Energy in Solar- and Battery-powered Computer Systems

联合管理太阳能和电池供电计算机系统中的电能和热能

• DOI: 10.1145/3575813.3595191
• 发表时间: 2023
• 期刊: 14th ACM International Conference on Future Energy Systems
• 作者: Bashir, Noman;Chandio, Yasra;Irwin, David;Anwar, Fatima M.;Gummeson, Jeremy;Shenoy, Prashant
• 通讯作者: Shenoy, Prashant

Model-driven Cluster Resource Management for AI Workloads in Edge Clouds

• DOI: 10.1145/3582080
• 发表时间: 2022-01
• 期刊: ACM Transactions on Autonomous and Adaptive Systems
• 影响因子: 2.7
• 作者: Qianlin Liang;Walid A. Hanafy;Ahmed Ali-Eldin;Prashant Shenoy

DACF: day-ahead carbon intensity forecasting of power grids using machine learning

DACF：使用机器学习对电网日前碳强度进行预测

• DOI: 10.1145/3538637.3538849
• 发表时间: 2022
• 期刊: Proceedings of the Thirteenth ACM International Conference on Future Energy Systems (e-Energy ’22
• 作者: Maji, Diptyaroop;Sitaraman, Ramesh K.;Shenoy, Prashant
• 通讯作者: Shenoy, Prashant

Understanding the Benefits of Hardware-Accelerated Communication in Model-Serving Applications

了解模型服务应用程序中硬件加速通信的好处

• DOI: 10.1109/iwqos57198.2023.10188785
• 发表时间: 2023
• 期刊: Proceedings of IEEE/ACM 31st International Symposium on Quality of Service (IWQoS
• 作者: Hanafy, Walid A.;Wang, Limin;Chang, Hyunseok;Mukherjee, Sarit;Lakshman, T. V.;Shenoy, Prashant
• 通讯作者: Shenoy, Prashant