Resource Allocation Reinforcement Learning for Quality of Service Maintenance in Cloud-Based Services

被引:2
作者
Hong, Dupyo [1 ]
Kim, DongWan [1 ]
Min, Oh Jung [1 ]
Shin, Yongtae [1 ]
机构
[1] Soongsil Univ, Dept Comp, Seoul, South Korea
来源
2023 INTERNATIONAL CONFERENCE ON INFORMATION NETWORKING, ICOIN | 2023年
关键词
Resource Allocation; Reinforcement Learning; Deep Q Network; Cloud Computing;
D O I
10.1109/ICOIN56518.2023.10048905
中图分类号
TP [自动化技术、计算机技术];
学科分类号
0812 ;
摘要
Recently, in order to improve the service quality of cloud-based services, research on a reinforcement learning model that predicts an appropriate amount of cloud resources by identifying patterns of user demands is being conducted. Reinforcement learning Q-learning algorithms rely on building a table (Q-table) for Q values, so if the state space and action space are vastly larger, they do not obtain optimal policies. In addition, learning errors in false experiences from the correlation of successive data in reinforcement learning may exist. In this paper, we study reinforcement learning modeling techniques that achieve higher accuracy than existing models by reducing the state definition space of hardware resources arising from services. It is possible to maximize service quality by allocating cloud resources or returning unnecessary resources with accurate resource demand prediction. For performance analysis, the service request prediction results according to the number of learnings were confirmed, and the service request prediction accuracy of three different models according to the neural network was compared. In the experiment, the model applying the proposed Convolutional Neural Network(CNN) neural network modeling technique is found to predict the amount of cloud resources in close proximity to the actual service request as the number of learning increases. We also compare the average of service request prediction accuracy of different models applying three neural networks, Deep Neural Network(DNN), Long Short-Term Memory(LTSM), and CNN, and find that the proposed technique has 3.36% higher prediction accuracy than LSTM-based models, and 40.2% higher than DNN-based models. In the future, additional research is needed, such as building various learning datasets or applying other reinforcement learning algorithms. Further research is also needed on cloud resource rental costs and provisioning latency.
引用
收藏
页码:517 / 521
页数:5
相关论文
共 13 条
[1]  
Alwarafy Abdulmalik., 2021, Deep reinforcement learning for radio resource allocation and management in next generation heterogeneous wireless networks: A survey
[2]   Resource Allocation for Cloud-Based Software Services Using Prediction-Enabled Feedback Control With Reinforcement Learning [J].
Chen, Xing ;
Zhu, Fangning ;
Chen, Zheyi ;
Min, Geyong ;
Zheng, Xianghan ;
Rong, Chunming .
IEEE TRANSACTIONS ON CLOUD COMPUTING, 2022, 10 (02) :1117-1129
[3]  
김영우, 2017, [Journal of the Institute of Electronics and Information Engineers, 전자공학회논문지], V54, P113
[4]   Reinforcement Learning Based Resource Management for Network Slicing [J].
Kim, Yohan ;
Kim, Sunyong ;
Lim, Hyuk .
APPLIED SCIENCES-BASEL, 2019, 9 (11)
[5]  
Lee S., 2020, P KIIT C, P472
[6]  
Lee Yeseul, 2022, [Journal of Digital Contents Society, 디지털콘텐츠학회논문지], V23, P531, DOI 10.9728/dcs.2022.23.3.531
[7]   Cloud spot instance price prediction using k NN regression [J].
Liu, Wenqiang ;
Wang, Pengwei ;
Meng, Ying ;
Zhao, Caihui ;
Zhang, Zhaohui .
HUMAN-CENTRIC COMPUTING AND INFORMATION SCIENCES, 2020, 10 (01)
[8]   Resource Management in Cloud Networking Using Economic Analysis and Pricing Models: A Survey [J].
Luong, Nguyen Cong ;
Wang, Ping ;
Niyato, Dusit ;
Wen, Yonggang ;
Han, Zhu .
IEEE COMMUNICATIONS SURVEYS AND TUTORIALS, 2017, 19 (02) :954-1001
[9]   New scheduling approach using reinforcement learning for heterogeneous distributed systems [J].
Orhean, Alexandru Iulian ;
Pop, Florin ;
Raicu, Ioan .
JOURNAL OF PARALLEL AND DISTRIBUTED COMPUTING, 2018, 117 :292-302
[10]  
Precedence Research, 2022, CLOUD SERV MARK SIZ