One Can Only Gain by Replacing EASY Backfilling: A Simple Scheduling Policies Case Study

Abstract

High-Performance Computing (HPC) platforms are growing in size and complexity. In order to improve the quality of service of such platforms, researchers are devoting a great amount of effort to devise algorithms and techniques to improve different aspects of performance such as energy consumption, total usage of the platform, and fairness between users. In spite of this, system administrators are always reluctant to deploy state of the art scheduling methods and most of them revert to EASY-backfilling, also known as EASY-FCFS (EASY-First-Come-First-Served). Newer methods frequently are complex and obscure and the simplicity and transparency of EASY are too important to sacrifice. In this work, we used execution logs from five HPC platforms to compare four simple scheduling policies: FCFS, Shortest estimated Processing time First (SPF), Smallest Requested Resources First (SQF), and Smallest estimated Area First (SAF). Using simulations, we performed a thorough analysis of the cumulative results for up to 180 weeks and considered three scheduling objectives: waiting time, slowdown and per-processor slowdown. We also evaluated other effects, such as the relationship between job size and slowdown, the distribution of slowdown values, and the number of backfilled jobs, for each HPC platform and scheduling policy. We conclude that one can only gain by replacing EASYbackfilling with SAF with backfilling, as it offers improvements in performance by up to 80% in the slowdown metric while maintaining the simplicity and the transparency of FCFS. Moreover, SAF reduces the number of jobs with large slowdowns and the inclusion of a simple thresholding mechanism guarantees that no starvation occurs. Finally, we propose SAF as a new benchmark for future scheduling studies.

RLSchert: An HPC Job Scheduler Using Deep Reinforcement Learning and Remaining Time Prediction

“…The completion time C j represents the response time of job j. The reward function of the completion time is defined in Equation (10). The sum of R t at each time step t is equal to the sum of the completion time.…”

“…In HPC, job scheduling has been a long-standing research topic [2][3][4][5][6][7][8][9][10][11][12]29]. Maximizing resource utilization, reducing resource fragmentation, and improving user satisfaction have always been the goals of researchers.…”

“…Maximizing resource utilization, reducing resource fragmentation, and improving user satisfaction have always been the goals of researchers. Therefore, various scheduling strategies have been proposed and developed, ranging from the simplest and straightforward heuristic algorithms (such as first come, first served (FCFS), short job first (SJF) and short area first (SAF) [10]) to more sophisticated policies (WFP3 and UNICEF) [4]. Subsequently, machine learning techniques have been applied in the field of scheduling.…”

“…As a wellknown NP-complete problem [3], the most straightforward and common solution is to make scheduling decisions through a heuristic priority function, which assigns the priority to each job based on its attributes [3][4][5][6][7][8]. Some scheduling strategies use the simplest priority function, such as first come, first served (FCFS) using job submission time, shortest job first (SJF) [9] using runtime, and smallest area first (SAF) [10] using the job requested area: runtime × resources. In addition, some schedulers calculate the priority based on more attributes such as UNICEF [4] and F1 [5].…”

Improving the performance of batch schedulers using online job runtime classification

“…One insight one can have is that the key factor of heuristics that favor shorter jobs, such as SPF, is that executing small jobs first improves the metrics, such as the job flow time and the slowdown [13]. In this work, we follow this insight and we propose to apply a simple two-class classification instead of regression.…”