Skip to Main content Skip to Navigation

Learning about simple heuristics for online parallel job scheduling

Abstract : High-Performance Computing (HPC) platforms are growing in size and complexity. In an adversarial manner, the power demand of such platforms has rapidly grown as well, and current top supercomputers require power at the scale of an entire power plant. In an effort to make a more responsible usage of such power, researchers are devoting a great amount of effort to devise algorithms and techniques to improve different aspects of performance such as scheduling and resource management. But HPC platform maintainers are still reluctant to deploy state of the art scheduling methods and most of them revert to simple heuristics such as EASY Backfilling, which is based in a naive First-Come-First-Served (FCFS) ordering. Newer methods are often complex and obscure, and the simplicity and transparency of EASY Backfilling are too important to sacrifice.At a first moment we explored Machine Learning (ML) techniques to learn on-line parallel job scheduling heuristics. Using simulations and a workload generation model, we could determine the characteristics of HPC applications (jobs) that lead to a reduction in the mean slowdown of jobs in an execution queue. Modeling these characteristics using a nonlinear function and applying this function to select the next job to execute in a queue improved the mean task slowdown in synthetic workloads. When applied to real workload traces from highly different machines, these functions still resulted in performance improvements, attesting the generalization capability of the obtained heuristics.At a second moment, using simulations and workload traces from several real HPC platforms, we performed a thorough analysis of the cumulative results of four simple scheduling heuristics (including EASY Backfilling). We also evaluated 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 show experimental evidence that one can only gain by replacing EASY Backfilling with the Smallest estimated Area First (SAF) policy with backfilling, as it offers improvements in performance by up to 80% in the slowdown metric while maintaining the simplicity and the transparency of EASY. SAF reduces the number of jobs with large slowdowns and the inclusion of a simple thresholding mechanism guarantees that no starvation occurs.Overall we achieved the following remarks: (i) simple and efficient scheduling heuristics in the form of a nonlinear function of the jobs characteristics can be learned automatically, though whether the reasoning behind their scheduling decisions is clear or not can be up to argument. (ii) The area (processing time estimate multiplied by the number of processors) of the jobs seems to be a quite important property for good parallel job scheduling heuristics, since many of the heuristics (notably SAF) that achieved good performances have the job's area as input. (iii) The backfilling mechanism seems to always help in increasing performance, though it does not outperform a better sorting of the jobs waiting queue, such as the sorting performed by SAF.
Complete list of metadatas

Cited literature [90 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, September 2, 2020 - 11:45:08 AM
Last modification on : Wednesday, October 14, 2020 - 4:18:51 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02928077, version 1



Danilo Carastan dos Santos. Learning about simple heuristics for online parallel job scheduling. Distributed, Parallel, and Cluster Computing [cs.DC]. Université Grenoble Alpes; Universidade Federal do ABC, 2019. English. ⟨NNT : 2019GREAM052⟩. ⟨tel-02928077⟩



Record views


Files downloads