Collaborative Research: CNS Core: Small: HARMONIA: New Methods for Colocating Multiple QoS-Sensitive Jobs

数据中心和高性能计算（HPC）系统被认为是所有现代计算需求的支柱，从Web搜索到电子邮件;从视频流到文件共享;从社交媒体平台到科学计算。如今，当代数据中心作业管理者在同一物理服务器上共同运行的应用程序之间采用保守的资源共享策略。当前的策略是保守的，以确保满足延迟关键型应用程序的严格延迟要求;然而，这种保守性导致昂贵的计算资源的巨大利用不足，这会产生资本和运营费用。HARMONIA提出了一系列利用贝叶斯优化（BO）原理的新型非常规资源策略，但为BO引入了新的创新，并证明了其对数据中心资源管理的有用性。HARMONIA将使用基于BO的学习模型捕获资源分配对应用程序性能的影响，并对共享资源进行分区，并相应地调整硬件/软件旋钮，以最大限度地提高单个应用程序的性能和系统利用率。为了实现实用性和可扩展性，HARMONIA采用了一组近似准确的在线学习模型，这些模型是轻量级的，而不是重量级的、完全准确的模型。传入的应用程序通过HARMONIA运行时框架以动态、高效和非侵入的方式与现有应用程序放在一起。该项目的成果将影响和冲击现代数据中心的运营，这些数据中心服务于我们对各种工作负载的计算需求，包括短期运行的延迟关键型应用程序（例如，机器学习推理、web搜索查询、微服务）和长时间运行的面向吞吐量的工作负载（例如，科学模拟）。提高大规模数据中心和HPC系统的利用率将带来更好的成本节约和更低的碳足迹。HARMONIA项目计划开展的教育和外联活动包括加强研究生课程，并推出关于“计算机系统概念”的新的每月播客，以更好地吸引高中生并使他们做好准备。所有开发的工具，软件工件，测量的数据集将提供给研究界，以进一步提高项目的成果和他们的影响。这个奖项反映了NSF的法定使命，并已被认为是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的支持。

Data centers and high performance computing (HPC) systems are considered the backbone of all modern-day computational needs for services ranging from Web search to emails; from video streaming to file sharing; from social media platforms to scientific computing. Today, contemporary data center job schedulers employ conservative resource sharing strategies among applications co-running on the same physical server. Current strategies are conservative to ensure that tight latency requirements for latency-critical applications are met; however, this conservatism leads to huge underutilization of expensive computing resources, which incurs both capital and operational expenses. HARMONIA proposes a family of novel unconventional resource strategies leveraging the principles of Bayesian Optimization (BO), but introducing novel innovations to BO and demonstrating its usefulness toward data center resource management. HARMONIA will capture the impact of resource allocation on application performance using BO-based learning models, and partition the shared resources and adjust hardware/software knobs accordingly to maximize the performance of individual applications and the system utilization. To achieve practicality and scalability, HARMONIA employs a pool of approximately-accurate online learning models which are lightweight instead of a heavyweight, fully-accurate model. Incoming applications are placed and co-located with existing applications in a dynamic, efficient, and non-intrusive manner by the HARMONIA runtime framework. Outcomes of this project will influence and impact the operations of modern data centers, which serve our computational needs for a variety of workloads including short-running latency-critical application (e.g., machine learning inferences, web search queries, microservices) and long-running throughput-oriented workloads (e.g., scientific simulations). Improving the utilization of large-scale data centers and HPC systems will lead to better cost savings and a lower carbon footprint. Planned educational and outreach activities for the project HARMONIA include enhancing graduate coursework and introducing a new monthly podcast on ``concepts in computer systems'' to better engage and prepare high school students. All developed tools, software artifacts, measured datasets will be made available to the research community for further enhancing the project outcomes and their impact.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.