M-DRL: Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption

Tianba Chen; Wei Li; Yu Kang Sun; Yunchun Li

doi:10.1007/978-3-030-79478-1_7

M-DRL: Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption

Tianba Chen
, Wei Li^*
, Yu Kang Sun
, Yunchun Li

^*此作品的通讯作者

计算机学院

Beihang University

科研成果: 书/报告/会议事项章节 › 会议稿件 › 同行评审

摘要

The coflow scheduling in data-parallel clusters can improve application-level communication performance. The existing coflow scheduling method without prior knowledge usually uses Multi-Level Feedback Queue (MLFQ) with fixed threshold parameters, which is insensitive to coflow traffic characteristics. Manual adjustment of the threshold parameters for different application scenarios often has long optimization period and is coarse in optimization granularity. We propose M-DRL, a deep reinforcement learning based coflow traffic scheduler by dynamically setting thresholds of MLFQ to adapt to the coflow traffic characteristics, and reduces the average coflow completion time. Trace-driven simulations on the public dataset show that coflow communication stages using M-DRL complete 2.08 × (6.48 × ) and 1.36 × (1.25 × ) faster on average coflow completion time (95-th percentile) in comparison to per-flow fairness and Aalo, and is comparable to SEBF with prior knowledge.

源语言	英语
主期刊名	Network and Parallel Computing - 17th IFIP WG 10.3 International Conference, NPC 2020, Revised Selected Papers
编辑	Xin He, En Shao, Guangming Tan
出版商	Springer Science and Business Media Deutschland GmbH
页	80-91
页数	12
ISBN（印刷版）	9783030794774
DOI	https://doi.org/10.1007/978-3-030-79478-1_7
出版状态	已出版 - 2021
活动	17th IFIP WG 10.3 International Conference on Network and Parallel Computing, NPC 2020 - Zhengzhou, 中国期限: 28 9月 2020 → 30 9月 2020

出版系列

姓名	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
卷	12639 LNCS
ISSN（印刷版）	0302-9743
ISSN（电子版）	1611-3349

会议

会议	17th IFIP WG 10.3 International Conference on Network and Parallel Computing, NPC 2020
国家/地区	中国
市	Zhengzhou
时期	28/09/20 → 30/09/20

访问文件

10.1007/978-3-030-79478-1_7

其它文件与链接

链接到 Scopus 的出版物

引用此

Chen, T., Li, W., Sun, Y. K., & Li, Y. (2021). M-DRL: Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption. 在 X. He, E. Shao, & G. Tan (编辑), Network and Parallel Computing - 17th IFIP WG 10.3 International Conference, NPC 2020, Revised Selected Papers (页码 80-91). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); 卷 12639 LNCS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-79478-1_7

Chen, Tianba ; Li, Wei ; Sun, Yu Kang 等. / M-DRL : Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption. Network and Parallel Computing - 17th IFIP WG 10.3 International Conference, NPC 2020, Revised Selected Papers. 编辑 / Xin He ; En Shao ; Guangming Tan. Springer Science and Business Media Deutschland GmbH, 2021. 页码 80-91 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{50824890ce474c389b07ccc2596c373c,

title = "M-DRL: Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption",

abstract = "The coflow scheduling in data-parallel clusters can improve application-level communication performance. The existing coflow scheduling method without prior knowledge usually uses Multi-Level Feedback Queue (MLFQ) with fixed threshold parameters, which is insensitive to coflow traffic characteristics. Manual adjustment of the threshold parameters for different application scenarios often has long optimization period and is coarse in optimization granularity. We propose M-DRL, a deep reinforcement learning based coflow traffic scheduler by dynamically setting thresholds of MLFQ to adapt to the coflow traffic characteristics, and reduces the average coflow completion time. Trace-driven simulations on the public dataset show that coflow communication stages using M-DRL complete 2.08 × (6.48 × ) and 1.36 × (1.25 × ) faster on average coflow completion time (95-th percentile) in comparison to per-flow fairness and Aalo, and is comparable to SEBF with prior knowledge.",

keywords = "Coflow, Datacenter network, Deep reinforcement learning",

author = "Tianba Chen and Wei Li and Sun, \{Yu Kang\} and Yunchun Li",

note = "Publisher Copyright: {\textcopyright} 2021, IFIP International Federation for Information Processing.; 17th IFIP WG 10.3 International Conference on Network and Parallel Computing, NPC 2020 ; Conference date: 28-09-2020 Through 30-09-2020",

year = "2021",

doi = "10.1007/978-3-030-79478-1\_7",

language = "英语",

isbn = "9783030794774",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "80--91",

editor = "Xin He and En Shao and Guangming Tan",

booktitle = "Network and Parallel Computing - 17th IFIP WG 10.3 International Conference, NPC 2020, Revised Selected Papers",

address = "德国",

}

Chen, T, Li, W, Sun, YK & Li, Y 2021, M-DRL: Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption. 在 X He, E Shao & G Tan (编辑), Network and Parallel Computing - 17th IFIP WG 10.3 International Conference, NPC 2020, Revised Selected Papers. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 卷 12639 LNCS, Springer Science and Business Media Deutschland GmbH, 页码 80-91, 17th IFIP WG 10.3 International Conference on Network and Parallel Computing, NPC 2020, Zhengzhou, 中国, 28/09/20. https://doi.org/10.1007/978-3-030-79478-1_7

M-DRL: Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption. / Chen, Tianba; Li, Wei; Sun, Yu Kang 等.
Network and Parallel Computing - 17th IFIP WG 10.3 International Conference, NPC 2020, Revised Selected Papers. 编辑 / Xin He; En Shao; Guangming Tan. Springer Science and Business Media Deutschland GmbH, 2021. 页码 80-91 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); 卷 12639 LNCS).

科研成果: 书/报告/会议事项章节 › 会议稿件 › 同行评审

TY - GEN

T1 - M-DRL

T2 - 17th IFIP WG 10.3 International Conference on Network and Parallel Computing, NPC 2020

AU - Chen, Tianba

AU - Li, Wei

AU - Sun, Yu Kang

AU - Li, Yunchun

PY - 2021

Y1 - 2021

N2 - The coflow scheduling in data-parallel clusters can improve application-level communication performance. The existing coflow scheduling method without prior knowledge usually uses Multi-Level Feedback Queue (MLFQ) with fixed threshold parameters, which is insensitive to coflow traffic characteristics. Manual adjustment of the threshold parameters for different application scenarios often has long optimization period and is coarse in optimization granularity. We propose M-DRL, a deep reinforcement learning based coflow traffic scheduler by dynamically setting thresholds of MLFQ to adapt to the coflow traffic characteristics, and reduces the average coflow completion time. Trace-driven simulations on the public dataset show that coflow communication stages using M-DRL complete 2.08 × (6.48 × ) and 1.36 × (1.25 × ) faster on average coflow completion time (95-th percentile) in comparison to per-flow fairness and Aalo, and is comparable to SEBF with prior knowledge.

AB - The coflow scheduling in data-parallel clusters can improve application-level communication performance. The existing coflow scheduling method without prior knowledge usually uses Multi-Level Feedback Queue (MLFQ) with fixed threshold parameters, which is insensitive to coflow traffic characteristics. Manual adjustment of the threshold parameters for different application scenarios often has long optimization period and is coarse in optimization granularity. We propose M-DRL, a deep reinforcement learning based coflow traffic scheduler by dynamically setting thresholds of MLFQ to adapt to the coflow traffic characteristics, and reduces the average coflow completion time. Trace-driven simulations on the public dataset show that coflow communication stages using M-DRL complete 2.08 × (6.48 × ) and 1.36 × (1.25 × ) faster on average coflow completion time (95-th percentile) in comparison to per-flow fairness and Aalo, and is comparable to SEBF with prior knowledge.

KW - Coflow

KW - Datacenter network

KW - Deep reinforcement learning

UR - https://www.scopus.com/pages/publications/85112633785

U2 - 10.1007/978-3-030-79478-1_7

DO - 10.1007/978-3-030-79478-1_7

M3 - 会议稿件

AN - SCOPUS:85112633785

SN - 9783030794774

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 80

EP - 91

BT - Network and Parallel Computing - 17th IFIP WG 10.3 International Conference, NPC 2020, Revised Selected Papers

A2 - He, Xin

A2 - Shao, En

A2 - Tan, Guangming

PB - Springer Science and Business Media Deutschland GmbH

Y2 - 28 September 2020 through 30 September 2020

ER -

Chen T, Li W, Sun YK, Li Y. M-DRL: Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption. 在 He X, Shao E, Tan G, 编辑, Network and Parallel Computing - 17th IFIP WG 10.3 International Conference, NPC 2020, Revised Selected Papers. Springer Science and Business Media Deutschland GmbH. 2021. 页码 80-91. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-030-79478-1_7

M-DRL: Deep Reinforcement Learning Based Coflow Traffic Scheduler with MLFQ Threshold Adaption

摘要

出版系列

会议

访问文件

其它文件与链接

指纹

引用此