Topic-Aware Load Balancing and Optimal Caching in Event-Driven Multi-Broker Networks for Edge-Assisted IoT

Abstract

With the rise of modern networking approaches, including Publish-Subscribe (Pub-Sub) networks and Edge-supported Internet of Things (IoT) systems, a new era of managing data and optimizing networks has begun. Pub-Sub networks decouple the producers and consumers of information, allowing for more scalable and flexible communication by enabling asynchronous data exchange through intermediaries called brokers. These fresh setups bring challenges and opportunities to improve how efficiently data is handled and how networks scale. This thesis focuses on enhancing networks, aiming to make Pub-Sub systems smoother and more efficient. The goal is to refine how data is managed and improve network performance in dynamic, widely spread-out environments. In Pub-Sub networks, brokers are intermediaries facilitating data delivery based on subscribed topics. However, the geographic distribution of brokers in multi-broker environments often leads to service disparities and intolerable delivery delays during peak events. To address this gap, we propose a topic-aware low-latency load balancing approach that leverages Trie data structures for topic association, outlier detection for identifying hot topics, and sojourn time modeling for load balancing. Experimental validation using a Pub-Sub MQTT testbed significantly improves average client waiting time, load distribution, and server utilization rates. Furthermore, caching optimization is a pivotal solution for reducing latency and congestion in Edge-assisted IoT frameworks. While Edge caching offers promising solutions, data transiency and network volatility present significant obstacles. We propose a Deep Reinforcement Learning (DRL) approach for caching action optimization, leveraging a Distributed Proximal Policy Optimization (DPPO) algorithm. By framing the caching problem as a Markov Decision Process (MDP), we balance data transiency and communication costs to minimize data retrieval costs. Our evaluation using RLlib in a distributed Pub-Sub network shows significant improvements in cache hit rates and faster convergence compared to existing DRL and baseline solutions. Despite these promising results, scalability in handling transient IoT data remains a challenge. The existing approach has limitations, such as not considering the lifetime attribute during model training and relying on a single agent to capture the dynamics of transient data. To address these issues, we advocate for decentralized and adaptive caching through localized training in multiple Edge devices. Leveraging Multi-Agent Deep Reinforcement Learning (MADRL), specifically the Multi-Agent Advantage Actor-Critic (MAA2C) algorithm, our IoT-C2 cooperative caching approach surpasses baseline policies in terms of data placement, freshness, delivery delay, and broker utilization. Optimizing energy consumption is essential in Edge-assisted IoT networks, where sensors play a crucial role in collecting and transmitting data. By caching frequently accessed data locally, sensors conserve energy, reducing the need to transmit data over long distances or through multiple network hops. We introduce an optimal custom caching strategy formulated as an MDP to jointly optimize long-term cumulative costs, considering user Age of Information (AoI) averages, sensor batteries, and data rates. Simulation results highlight the effectiveness of our proposed strategy in reducing AoI compared to conventional approaches, underscoring the importance of customized caching for Edge computing. This study addresses critical challenges in modern networking paradigms, focusing on efficient data management strategies across diverse environments such as Pub-Sub networks and Edge-assisted IoT frameworks. By delving into topics like load balancing in Pub-Sub networks and caching optimization in Edge-assisted IoT systems, it aims to optimize data delivery, reduce latency, and improve overall network performance. Through innovative methodologies and algorithms, this research endeavors to enhance system scalability, minimize delivery delays, and optimize resource utilization, facilitating seamless data transmission and enhancing the reliability and efficiency of networked systems.