EAGER: From Theory to Practice of Elastic Interval Runtime Schedulers

点击翻译按钮获取中文摘要

Data centers and high-performance computing systems (HPC) across the globe form the backbone of modern-day computing services and applications, such as Internet searches, scientific discoveries, emails, weather predictions, social media platforms, and data sharing. Resource (computational, memory, and network) orchestration within data centers and HPC, is, and always has been, a challenge, the foremost reason being that application requirements and resource utilization cannot be predicted with 100% accuracy. Consequently, resource-orchestration methodologies undertake critical assumption on the application needs (for example, job arrival rates and job completions times) to utilize resources. More than often, such assumptions lead to bottlenecks or underutilization. This project investigates a new class of resource-orchestration methodologies, called the Elastic Interval Runtime Schedulers (EIRIS), that do not undertake assumptions on the job arrival rates or completion times or bounds on the application parallelism. The project will develop practical solutions for EIRIS anchored in solid mathematical underpinnings. The project is expected to have a profound impact on the scheduling theory and runtime systems for data centers and HPC. Applications immediately benefiting from this work include (but are not limited to): cloud gaming, video coding, and video transcoding. The project research findings will be integrated into a graduate class on computer systems. Moreover, a monthly podcast, entitled concepts in computer systems, will be broadcasted to Mississippi high school students with the goal to encourage them to pursue education in computer and information science and engineering.The project will develop the mathematical underpinnings, theoretical bounds, and translatable runtime for EIRIS. The project will be divided into three interleaving thrusts focusing on: (1) mathematical underpinnings for EIRIS by utilizing concepts and operations from group theory to: (a) develop an oracle for analyses, (b) derive theoretical lower bounds, and (c) make an attempt to further refine the lower bounds; (2) developing a set of resource-orchestration methodologies for EIRIS by utilizing concepts from bin packing algorithms to develop procedures predominantly relying on: (a) actual job duration times and (b) remaining job duration times; and (3) an autonomic runtime for EIRIS in G-Hadoop by implementing (a) distributed resource management systems, (b) metadata servers, and (c) packages for reproducible artifacts. All thrusts will be deployed and tested on large-scale computer systems for demonstration.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.