Towards an Optimized Distributed Message Queue System for AIoT Edge Computing: A Reinforcement Learning Approach
Abstract
:1. Introduction
 We propose a distributed message system for largescale AIoT based on Kafka to address message ordering challenges in AIoT edge computing. The impact of different factors on system performance in distributed AIoT messaging scenarios is investigated. A partition selection algorithm (PSA) is specifically designed for the proposed distributed message queues, aiming to maintain the order of AIoT messages, balance the load among broker clusters, and enhance availability during the publication of subscribable messages by AIoT edge devices.
 We propose a reinforcementlearningbased method called DMSCO (DDPGbased distributed message queue systems configuration optimization) that utilizes a preprocessed parameter list as an action space to train our decision model. By incorporating rewards based on the distributed message queue system’s throughput and message transmission success rate, DMSCO efficiently optimizes messaging performance in AIoT scenarios by adaptively finetuning parameter configurations.
 We conducted a comprehensive evaluation of the proposed DMSCO algorithm, assessing its performance efficacy for the distributed message queue system in AIoT edge computing scenarios across varying message sizes and transmission frequencies. Through comparative analysis against methods employing genetic algorithms and random searching, we observed that the DMSCO algorithm provides an improved solution to meet the specific demands of largerscale, highconcurrency AIoT edge computing applications.
2. Related Work
3. Distributed Message Queue System for AIoT Edge Computing
3.1. Distributed Message System for LargeScale AIoT Edge Computing
Algorithm 1: Partition selection algorithm (PSA). 
3.2. Performance Modeling in AIoT Edge Computing Scenarios
4. ReinforcementLearningBased Method for Optimized AIoT Message Queue System
4.1. Parameter Screening
Algorithm 2: Dimensionality reduction method based on PCA for the initial training sample set. 
Input: Original samples $X=\{{X}_{1},{X}_{2},{X}_{3},\cdots ,{X}_{22}\}$, where each row represents values of each parameter in the training samples and each column represents the data of the ith sample Output: The final sample dataset Y

4.2. LassoRegressionBased Performance Modeling
Algorithm 3: Performance modeling and key parameters screening by Lasso regression. 
Input: Preprocessed samples $Y=\{{Y}_{1},{Y}_{2},{Y}_{3},\cdots ,{Y}_{22}\}$ Output: Key parameters and their weightings

4.3. Optimization Method Based on Deep Deterministic Policy Gradient Algorithm
Algorithm 4: DDPGbased distributed message system configuration optimization (DMSCO). 
 Environment: The environment refers to the distributed message system being optimized. We utilize the performance model built through the Lasso regression as the simulated edge environment, wherein the resultant increase or decrease in the system’s throughput serves as a performancebased reward.
 Agent: The configuration optimizer based on DDPG is regarded as the agent.
 Action: Action is depicted as a vector consisting of adjustable parameters.
 State: State can refer to the system running metrics.
 Reward: The reward is defined as the augmentation in throughput relative to both the initial configuration and the preceding one.
4.4. Complexity Analysis
 Forward propagation through the actor and critic networks occurs at each time step t within the range $1,2,\dots ,T$. During these passes, we perform computations on the actor and critic networks. Assuming the complexity of the forward pass for the actor network is $O\left(A\right)$, and for the critic network is $O\left(C\right)$, the overall complexity for T time steps is $O(T\xb7(A+C\left)\right)$. It is important to note that the complexity of $O\left(A\right)$ or $O\left(C\right)$ depends on the specific size of the model, as these steps involve matrix multiplications.
 During each time step, a backward propagation is performed to calculate gradients for both the actor and critic networks. This step involves computing the gradients for the actor network with a complexity of $O\left({A}_{\mathrm{grad}}\right)$ and for the critic network with a complexity of $O\left({C}_{\mathrm{grad}}\right)$. Considering T time steps, the total complexity becomes $O(T\xb7({A}_{\mathrm{grad}}+{C}_{\mathrm{grad}}))$.
 Gradient descent optimization involves performing the optimization process for each batch of size B. Considering the complexity of the optimization step as $O\left(\mathrm{Opt}\right)$, the overall complexity for T time steps can be estimated as $O(T\xb7\mathrm{Opt}/B)$, as an optimization step is executed for every B time steps.
 The target networks are updated periodically every $sync$ time steps, which is a hyperparameter to control the frequency of synchronization. Assume the complexity of updating the target networks is $O\left(\mathrm{Upd}\right)$. This depends on the complexity of replication between two identical network matrices. The overall complexity for T time steps can be approximated as $O(T\xb7\mathrm{Upd}/\mathrm{sync})$, as a target network update is performed for every $sync$ time steps.
5. Experiments
Analysis on Performance and Results
6. Conclusions
Author Contributions
Funding
Institutional Review Board Statement
Informed Consent Statement
Data Availability Statement
Conflicts of Interest
References
 Peres, R.S.; Jia, X.; Lee, J.; Sun, K.; Colombo, A.W.; Barata, J. Industrial artificial intelligence in industry 4.0systematic review, challenges and outlook. IEEE Access 2020, 8, 220121–220139. [Google Scholar] [CrossRef]
 Ullah, Z.; AlTurjman, F.; Mostarda, L.; Gagliardi, R. Applications of artificial intelligence and machine learning in smart cities. Comput. Commun. 2020, 154, 313–323. [Google Scholar] [CrossRef]
 Zhu, S.; Ota, K.; Dong, M. EnergyEfficient Artificial Intelligence of Things With Intelligent Edge. IEEE Internet Things J. 2022, 9, 7525–7532. [Google Scholar] [CrossRef]
 Chang, Z.; Liu, S.; Xiong, X.; Cai, Z.; Tu, G. A Survey of Recent Advances in EdgeComputingPowered Artificial Intelligence of Things. IEEE Internet Things J. 2021, 8, 13849–13875. [Google Scholar] [CrossRef]
 de Freitas, M.P.; Piai, V.A.; Farias, R.H.; Fernandes, A.M.R.; de Moraes Rossetto, A.G.; Leithardt, V.R.Q. Artificial Intelligence of Things Applied to Assistive Technology: A Systematic Literature Review. Sensors 2022, 22, 8531. [Google Scholar] [CrossRef] [PubMed]
 Baker, S.; Xiang, W. Artificial Intelligence of Things for Smarter Healthcare: A Survey of Advancements, Challenges, and Opportunities. IEEE Commun. Surv. Tutorials 2023, 25, 1261–1293. [Google Scholar] [CrossRef]
 Snyder, B.; Bosanac, D.; Davies, R. Introduction to Apache ActiveMQ. In Active MQ in Action; Manning Publications Co.: Shelter Island, NY, USA, 2011; pp. 6–16. [Google Scholar]
 Dinculeană, D.; Cheng, X. Vulnerabilities and limitations of MQTT protocol used between IoT devices. Appl. Sci. 2019, 9, 848. [Google Scholar] [CrossRef] [Green Version]
 Wu, H.; Shang, Z.; Wolter, K. Performance Prediction for the Apache Kafka Messaging System. In Proceedings of the 21st IEEE International Conference on High Performance Computing and Communications, Zhangjiajie, China, 10–12 August 2019; pp. 154–161. [Google Scholar]
 Li, R.; Yin, J.; Zhu, H. Modeling and Analysis of RabbitMQ Using UPPAAL. In Proceedings of the 2020 IEEE 19th International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom), Guangzhou, China, 29 December–1 January 2020; pp. 79–86. [Google Scholar] [CrossRef]
 Fu, G.; Zhang, Y.; Yu, G. A Fair Comparison of Message Queuing Systems. IEEE Access 2021, 9, 421–432. [Google Scholar] [CrossRef]
 Camposo, G. Messaging with Apache Kafka. In Cloud Native Integration with Apache Camel: Building Agile and Scalable Integrations for Kubernetes Platforms; Apress: Berkeley, CA, USA, 2021; pp. 167–209. [Google Scholar] [CrossRef]
 Johansson, L.; Dossot, D. RabbitMQ Essentials: Build Distributed and Scalable Applications with Message Queuing Using RabbitMQ; Packt Publishing Ltd.: Birmingham, UK, 2020. [Google Scholar]
 Leang, B.; Ean, S.; Ryu, G.A.; Yoo, K.H. Improvement of Kafka Streaming Using Partition and MultiThreading in Big Data Environment. Sensors 2019, 19, 134. [Google Scholar] [CrossRef] [Green Version]
 Wang, G.; Chen, L.; Dikshit, A.; Gustafson, J.; Chen, B.; Sax, M.J.; Roesler, J.; BleeGoldman, S.; Cadonna, B.; Mehta, A.; et al. Consistency and Completeness: Rethinking Distributed Stream Processing in Apache Kafka. In Proceedings of the 2021 International Conference on Management of Data (SIGMOD ’21), New York, NY, USA, 20–25 June 2021; pp. 2602–2613. [Google Scholar] [CrossRef]
 Jolliffe, I. A 50year personal journey through time with principal component analysis. J. Multivar. Anal. 2022, 188, 104820. [Google Scholar] [CrossRef]
 Wang, F.; Mukherjee, S.; Richardson, S.; Hill, S.M. Highdimensional regression in practice: An empirical study of finitesample prediction, variable selection and ranking. Stat. Comput. 2020, 30, 697–719. [Google Scholar] [CrossRef] [PubMed] [Green Version]
 Dilek, S.; Irgan, K.; Guzel, M.; Ozdemir, S.; Baydere, S.; Charnsripinyo, C. QoSaware IoT networks and protocols: A comprehensive survey. Int. J. Commun. Syst. 2022, 35, e5156. [Google Scholar] [CrossRef]
 Bayılmış, C.; Ebleme, M.A.; Ünal, Çavuşoğlu; Küçük, K.; Sevin, A. A survey on communication protocols and performance evaluations for Internet of Things. Digit. Commun. Netw. 2022, 8, 1094–1104. [Google Scholar] [CrossRef]
 Tariq, M.A.; Khan, M.; Raza Khan, M.T.; Kim, D. Enhancements and Challenges in CoAP—A Survey. Sensors 2020, 20, 6391. [Google Scholar] [CrossRef] [PubMed]
 da Cruz, M.A.; Rodrigues, J.J.; Lorenz, P.; Solic, P.; AlMuhtadi, J.; Albuquerque, V.H.C. A proposal for bridging application layer protocols to HTTP on IoT solutions. Future Gener. Comput. Syst. 2019, 97, 145–152. [Google Scholar] [CrossRef]
 Hesse, G.; Matthies, C.; Uflacker, M. How Fast Can We Insert? An Empirical Performance Evaluation of Apache Kafka. In Proceedings of the 2020 IEEE 26th International Conference on Parallel and Distributed Systems (ICPADS), Hong Kong, 2–4 December 2020; pp. 641–648. [Google Scholar] [CrossRef]
 Wu, H.; Shang, Z.; Wolter, K. Learning to Reliably Deliver Streaming Data with Apache Kafka. In Proceedings of the 2020 50th Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN), Valencia, Spain, 29 June–2 July 2020; pp. 564–571. [Google Scholar] [CrossRef]
 Donta, P.K.; Srirama, S.N.; Amgoth, T.; Annavarapu, C.S.R. Survey on recent advances in IoT application layer protocols and machine learning scope for research directions. Digit. Commun. Netw. 2022, 8, 727–744. [Google Scholar] [CrossRef]
 Dou, H.; Chen, P.; Zheng, Z. Hdconfigor: Automatically Tuning High Dimensional Configuration Parameters for Log Search Engines. IEEE Access 2020, 8, 80638–80653. [Google Scholar] [CrossRef]
 Ma, J.; Xie, S.; Zhao, J. NetMQ: Highperformance Innetwork Caching for Message Queues with Programmable Switches. In Proceedings of the IEEE International Conference on Communications, Seoul, Republic of Korea, 16–20 May 2022; pp. 4595–4600. [Google Scholar] [CrossRef]
 Dou, H.; Wang, Y.; Zhang, Y.; Chen, P. DeepCAT: A CostEfficient Online Configuration AutoTuning Approach for Big Data Frameworks. In Proceedings of the 51st International Conference on Parallel Processing (ICPP ’22), Bordeaux, France, 29 August–1 September 2022. [Google Scholar] [CrossRef]
 Dou, H.; Zhang, L.; Zhang, Y.; Chen, P.; Zheng, Z. TurBO: A costefficient configurationbased autotuning approach for clusterbased big data frameworks. J. Parallel Distrib. Comput. 2023, 177, 89–105. [Google Scholar] [CrossRef]
 Gou, F.; Wu, J. Message transmission strategy based on recurrent neural network and attention mechanism in IoT system. J. Circuits Syst. Comput. 2022, 31, 2250126. [Google Scholar] [CrossRef]
 Hong, L.; Deng, L.; Li, D.; Wang, H.H. Artificial intelligence pointtopoint signal communication network optimization based on ubiquitous clouds. Int. J. Commun. Syst. 2021, 34, e4507. [Google Scholar] [CrossRef]
 Lillicrap, T.P.; Hunt, J.J.; Pritzel, A.; Heess, N.; Erez, T.; Tassa, Y.; Silver, D.; Wierstra, D. Continuous control with deep reinforcement learning. In Proceedings of the 4th International Conference on Learning Representations, ICLR, San Juan, Puerto Rico, 2–4 May 2016. [Google Scholar]
 Cerda, P.; Varoquaux, G. Encoding HighCardinality String Categorical Variables. IEEE Trans. Knowl. Data Eng. 2022, 34, 1164–1176. [Google Scholar] [CrossRef]
 Katoch, S.; Chauhan, S.S.; Kumar, V. A review on genetic algorithm: Past, present, and future. Multimed. Tools Appl. 2021, 80, 8091–8126. [Google Scholar] [CrossRef] [PubMed]
Protocol  MQTT  CoAP  AMQP  HTTP 

Communication Model  Publish/Subscribe  Request/Response  Publish/Subscribe  Request/Response 
Lightweight  Yes  Yes  No  No 
Bandwidth Efficiency  High  High  Medium  Low 
Power Consumption  Low  Low  Medium  Medium 
RealTime Support  Limited  Limited  Yes  Limited 
Security  Supplementary measures  Builtin options  Advanced options  Builtin options 
Scalability  High  Medium  High  High 
Parameter Name  Weight 

bThreads  15.03 
cType  70.35 
nNThreads  23.74 
nIThreads  25.16 
mMBytes  60.35 
qM·Requests  124.32 
nRFetchers  $24.59$ 
sRBBytes  70.42 
sSBBytes  120.35 
sRMBytes  54.36 
acks  43.58 
bMemory  73.66 
bSize  $170.95$ 
lMs  34.32 
Component  Specification/Version 

Operating system  Ubuntu 20.04.1 LTS 
CPU  48 CPUs—Intel Xeon Gold 6126 @ 2.60 GHz 
Memory  187 GB 
Hard drive  8.2 TB 
LAN speed  10 GbE 
Docker version  23.0.4 
Framework  Springboot 2.7 
Kafka image  wurstmeister/kafka:2.122.4.0 
Kafka Java client  Producer and consumer 
Maven libraries  springbootstarterweb: v2.1.4, springkafka:v2.1.7, lombok: 0.322018.2 
Methods  Scenario  Throughput  Success Rate 

DMSCO  Smallsize msg and high frequency  88.79 MB/s  57.78% 
Largesize msg and low frequency  108.50 MB/s  68.45%  
Genetic algorithm  Smallsize msg and high frequency  80.52 MB/s  49.80% 
Largesize msg and low frequency  92.32 MB/s  61.15%  
Random searching  Smallsize msg and high frequency  73.99 MB/s  45.38% 
Largesize msg and low frequency  79.03 MB/s  61.15%  
No optimization  Smallsize msg and high frequency  60.56 MB/s  39.54% 
Largesize msg and low frequency  58.51 MB/s  33.67% 
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. 
© 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Xie, Z.; Ji, C.; Xu, L.; Xia, M.; Cao, H. Towards an Optimized Distributed Message Queue System for AIoT Edge Computing: A Reinforcement Learning Approach. Sensors 2023, 23, 5447. https://doi.org/10.3390/s23125447
Xie Z, Ji C, Xu L, Xia M, Cao H. Towards an Optimized Distributed Message Queue System for AIoT Edge Computing: A Reinforcement Learning Approach. Sensors. 2023; 23(12):5447. https://doi.org/10.3390/s23125447
Chicago/Turabian StyleXie, Zaipeng, Cheng Ji, Lifeng Xu, Mingyao Xia, and Hongli Cao. 2023. "Towards an Optimized Distributed Message Queue System for AIoT Edge Computing: A Reinforcement Learning Approach" Sensors 23, no. 12: 5447. https://doi.org/10.3390/s23125447