AITF: Applied Algorithmic Foundation for Scheduling Multiprogrammed Parallelizable Workloads

大多数现代计算系统都包含由许多应用程序共享的多处理器硬件。在这些系统中，决定如何在应用程序之间分配资源的良好调度算法对于确保良好的服务质量和有效利用系统资源至关重要。本项目将设计基本算法和调度算法的原型实现，为这些共享机器提供性能和资源利用率的保证。这项工作将通过提高从个人计算机到数据中心再到超级计算机的并行系统的性能来塑造关键计算基础设施的效率。所有结果，包括已发表的文章和软件工件，将通过全球网络向公众发布。这些学院将把研究与教育结合起来，将研究纳入学院的研究生课程，并培养博士、硕士和学士学位学生进行应用和理论并行计算研究。该项目将包括在理论基础的指导下设计实际有效的调度器。pi将通过考虑对系统设计者重要的各种调度目标，设计并行程序的多程序调度理论。特别是，该研究将集中于优化服务器、云和交互式系统中使用的延迟目标。pi还将探索可用于在实践中实施这些算法的有效机制，并对其设计进行经验验证。通过将理论分析与经验评估的反馈相结合，所提出的工作将获得见解，这将推动并行作业调度理论和实践的艺术状态。

Most modern computing systems contain multiprocessor hardware which is shared by many applications. In these systems, good scheduling algorithms that decide how to allocate resources among applications are crucial to ensure good quality of service and efficient use of system resources. This project will design foundational algorithms and prototype implementations of scheduling algorithms that provide guarantees of performance and resource utilization for these shared machines. This work will shape the efficiency of critical computing infrastructure by improving performance of parallel systems from personal computers to data centers to supercomputers. All results, including published articles and software artifacts, will be released to the public via world-wide web. The PIs will integrate research with education by incorporating this research into the PIs' graduate courses and training PhD, MS, and BS students in applied and theoretical parallel computing research.The project will involve designing practically efficient schedulers guided by theoretical foundations. The PIs will design a theory of multi-programmed scheduling for parallel programs by considering a variety of scheduling objectives important to system designers. In particular, the research will focus on optimizing latency objectives that are used in servers, clouds, and interactive systems. The PIs will also explore efficient mechanisms that can be used to implement these algorithms in practice and perform empirical validations of their designs. By combining theoretical analysis with feedback from empirical evaluations, the proposed work will gain insights that will advance the state of the art of both theory and practice of parallel job scheduling.