<  Back to the Polytechnique Montréal portal

A DAQM-Based Load Balancing Scheme for High Performance Computing Platforms

Kaijun Yang, Meng Li, Guchuan Zhu and Yvon Savaria

Article (2017)

[img]
Preview
Accepted Version
Terms of Use: All rights reserved.
Download (1MB)
Cite this document: Yang, K., Li, M., Zhu, G. & Savaria, Y. (2017). A DAQM-Based Load Balancing Scheme for High Performance Computing Platforms. IEEE Access, 5, p. 22504-22513. doi:10.1109/access.2017.2760251
Show abstract Hide abstract

Abstract

This paper addresses the load balancing problem, which is one of the key issues in high-performance computing (HPC) platforms. A novel method, called decentralized active queue management (DAQM), is proposed to provide a fair task distribution in a heterogeneous computing environment for HPC platforms. An implementation of the DAQM is presented, which consists of an ON-OFF queue control and a utility maximization-based coordination scheme. The stability of the queue control scheme and the convergence of the algorithm for utility maximization have been assessed by rigorous analysis. To demonstrate the performance of the developed queueing control system, numerical simulations are carried out and the obtained results confirm the efficiency and viability of the developed scheme.

Uncontrolled Keywords

Decentralized active queue management (DAQM),load balancing,high performance computing

Open Access document in PolyPublie
Subjects: 2500 Génie électrique et électronique > 2500 Génie électrique et électronique
Department: Département de génie électrique
Research Center: Non applicable
Funders: CRSNG/NSERC, Huawei Technologies Canada Co., Ltd.
Date Deposited: 21 Dec 2017 15:00
Last Modified: 24 Oct 2018 16:12
PolyPublie URL: https://publications.polymtl.ca/2842/
Document issued by the official publisher
Journal Title: IEEE Access (vol. 5)
Publisher: IEEE
Official URL: https://doi.org/10.1109/access.2017.2760251

Statistics

Total downloads

Downloads per month in the last year

Origin of downloads

Dimensions

Repository Staff Only