Experience and Practice of Batch Scheduling on Leadership Supercomputers at Argonne

Allcock, William; Rich, Paul M.; Fan, Yuping; Lan, Zhiling

doi:10.1007/978-3-319-77398-8_1

Cited by 22 publications

(14 citation statements)

References 5 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…A good solution is supposed to achieve good scheduling performance, such as low average job wait time and high system utilization, with the minimum scheduling overhead. Typical HPC systems tolerate 10-30 s of scheduling delay [17,18]. It is crucial to evaluate the performance and overhead of new RL scheduling methods before deployment.…”

Section: Impactmentioning

confidence: 99%

DRAS-CQSim: A reinforcement learning based framework for HPC cluster scheduling

Fan

Lan

2021

Software Impacts

Self Cite

View full text Add to dashboard Cite

For decades, system administrators have been striving to design and tune cluster scheduling policies to improve the performance of high performance computing (HPC) systems. However, the increasingly complex HPC systems combined with highly diverse workloads make such manual process challenging, time-consuming, and error-prone. We present a reinforcement learning based HPC scheduling framework named DRAS-CQSim to automatically learn optimal scheduling policy. DRAS-CQSim encapsulates simulation environments, agents, hyperparameter tuning options, and different reinforcement learning algorithms, which allows the system administrators to quickly obtain customized scheduling policies.

show abstract

Section: Impactmentioning

confidence: 99%

DRAS-CQSim: A reinforcement learning based framework for HPC cluster scheduling

Fan

Lan

2021

Software Impacts

Self Cite

View full text Add to dashboard Cite

show abstract

“…However, fairness and utilization are conflicting goals and aggressively using fair sharing can hurt cluster utilization [19]. In contrast, HPC systems are designed to run large jobs and prefer users running large jobs [10]. Therefore, fair sharing is not a concern in HPC scheduling.…”

Section: Multi-resource Schedulingmentioning

confidence: 99%

“…In the past, a number of scheduling policies have been proposed, and one of the widely used policies is FCFS, which sorts the jobs in the order of their arrivals. At ALCF, to support the mission of running large-scale capability jobs, a utility-based scheduling policy, named WFP, is deployed which periodically calculates a priority increment for each waiting job [10,42]. EASY backfilling is a commonly used strategy to enhance system utilization, where subsequent jobs are allowed to skip ahead under the condition that they do not delay the job at the head of the queue [30].…”

Section: Background and Related Work 21 Hpc Schedulingmentioning

confidence: 99%

Scheduling Beyond CPUs for HPC

Fan

Lan

Rich

et al. 2019

Proceedings of the 28th International Symposium on High-Performance Parallel and Distributed Computing

Self Cite

View full text Add to dashboard Cite

High performance computing (HPC) is undergoing significant changes. The emerging HPC applications comprise both compute-and dataintensive applications. To meet the intense I/O demand from emerging data-intensive applications, burst buffers are deployed in production systems. Existing HPC schedulers are mainly CPU-centric. The extreme heterogeneity of hardware devices, combined with workload changes, forces the schedulers to consider multiple resources (e.g., burst buffers) beyond CPUs, in decision making. In this study, we present a multi-resource scheduling scheme named BB-Sched that schedules user jobs based on not only their CPU requirements, but also other schedulable resources such as burst buffer. BBSched formulates the scheduling problem into a multi-objective optimization (MOO) problem and rapidly solves the problem using a multi-objective genetic algorithm. The multiple solutions generated by BBSched enables system managers to explore potential tradeoffs among various resources, and therefore obtains better utilization of all the resources. The trace-driven simulations with real system workloads demonstrate that BBSched improves scheduling performance by up to 41% compared to existing methods, indicating that explicitly optimizing multiple resources beyond CPUs is essential for HPC scheduling.

show abstract

“…В свою очередь, отделение Аргоннской национальной лаборатории Argonne Leadership Computing Facility (ALCF) [23] ставит перед собой следующую цель эксплуатации своих систем: утилизация вычислительного времени системы является важной целью в ALCF, но самая большая цель это позволить заданиям экстремального масштаба максимально быстро запускаться [24]. С 2013 по 2017 г. суперкомпьютер "Mira" дал доступ к своим ресурсам более 550 проектам и 1000 пользователям, более 280 000 задач было запущено на этой системе.…”

unclassified

“…С 2013 по 2017 г. суперкомпьютер "Mira" дал доступ к своим ресурсам более 550 проектам и 1000 пользователям, более 280 000 задач было запущено на этой системе. Для решения своей амбициозной цели ALCF отошли от традиционного подхода к планированию ресурсов кластера, они придумали свою "функцию утилизации" и реализовали специальную систему очередей, которая позволяет максимально эффективно решать поставленную задачу [24].…”

unclassified

Target optimization of a supercomputer task flow

Leonenkov

2019

Numerical Methods and Programming (Vychislitel'nye Metody I Programmirovanie)

View full text Add to dashboard Cite

Настоящая статья является результатом исследования потока задач суперкомпьютеров "Ломоносов" и "Ломоносов-2". Предложен подход к оценке эффективности функционирования суперкомпьютерной системы, основанный на ее базовых характеристиках. Введена новая функция потери качества планирования суперкомпьютера. Подход позволяет сравнивать разные суперкомпьютерные системы исходя из целей использования, которые стоят перед системными администраторами вычислительных центров. Описан опыт применения целевой оптимизации процессов планирования, основанной на предложенном подходе, для суперкомпьютеров "Ломоносов" и "Ломоносов-2". This paper is a result of studying the task flows observed on the Lomonosov and Lomonosov-2 supercomputers. A new approach to evaluating the performance of a supercomputer system based on its basic performance characteristics is proposed. A supercomputer's scheduling efficiency function is introduced for Lomonosov, Lomonosov-2 and other systems. The approach allows the system administrators to compare various supercomputer systems based on their usage aims. This paper describes the Moscow State University experience of applying the proposed approach to the optimization of Lomonosov and Lomonosov-2 scheduling resources.

show abstract

Experience and Practice of Batch Scheduling on Leadership Supercomputers at Argonne

Cited by 22 publications

References 5 publications

DRAS-CQSim: A reinforcement learning based framework for HPC cluster scheduling

DRAS-CQSim: A reinforcement learning based framework for HPC cluster scheduling

Scheduling Beyond CPUs for HPC

Target optimization of a supercomputer task flow

Contact Info

Product

Resources

About