KJCKorea
Journal Central

This paper compares and analyzes the real-time large-scale data processing performance of relational databases (RDB) and NoSQL databases based on Spring WebFlux and Spring MVC, providing criteria for selecting the optimal technology when designing real-time data processing systems. WebFlux, with its asynchronous non-blocking model, offers high concurrency and low response time, whereas MVC, based on a synchronous blocking I/O model, delivers stable performance in environments with relatively fewer concurrent connections. Additionally, NoSQL outperforms RDB in terms of flexibility and scalability, exhibiting less performance degradation as the number of concurrent users increases. Experimental results show that in a Spring WebFlux environment, NoSQL achieved an average response time approximately 50% shorter than RDB. Even in an MVC environment, NoSQL demonstrated about 40% faster response times compared to RDB. Notably, when the number of concurrent users exceeded 5,000, the combination of WebFlux and NoSQL yielded the best performance.

With recent advancements in energy harvesting technologies and the proliferation of AIoT devices capable of on-device learning, energy-harvesting AIoT systems have attracted significant attention. This study proposes an energy-adaptive approximate computing method designed to efficiently enhance federated learning performance while ensuring the stable operation of energy-constrained edge AIoT devices. Typically, there is a proportional relationship between learning performance and energy consumption; thus, performance trade-offs are inevitable under stringent energy conditions. Based on a solar energy harvesting model, the proposed scheme optimizes local training data sizes and parameter exchange volumes for each training round within limited energy budgets, performing approximate computing. This approach effectively minimizes device blackout durations, thereby improving the overall accuracy of federated learning.

This paper proposes a flow-level QoS (Quality of Service) routing framework that integrates AI-based traffic prediction to improve real-time network performance. The proposed method uses a LSTM (Long Short-Term Memory) neural network to estimate future traffic loads based on time-series data collected in a SDN (Software-Defined Networking) environment. Using the predicted traffic, the SDN controller dynamically adjusts link weights considering latency, congestion, and packet loss, and computes optimized paths via a modified Dijkstra algorithm. The system was implemented in a Mininet-based SDN testbed. Experimental results show that the proposed LSTM-based routing method significantly outperforms both static and non-predictive dynamic routing in terms of delay, packet loss, throughput, and path stability. Average latency was reduced by 30%, with routing decision time maintained below 100ms. Additionally, this system enables fast integration of predictions through a RESTful API that connects the predictor and SDN controller. In this paper, we experimentally demonstrate the effectiveness and adaptability of prediction-based QoS routing, and show that it is a structure that can be applied to future real-time network environments such as high-reliability networks, large-scale IoT environments, and smart city systems.