Future 6G use cases—such as AI, sensing, and immersive Extended Reality (XR)—will require stringent end-to-end communication performance. Optimizing only the network infrastructure is insufficient; coordination between network and computing resources is essential. In this paper, we propose a unified architecture that embeds a Layer 4 proxy into the User Plane Function (UPF) to control data flow and computing resources jointly. This design allows the flow to be dynamically managed based on network conditions, while simultaneously optimizing latency and throughput. The control interface is exposed via APIs, enabling applications to dynamically activate computing services and optimize data throughput for available network bandwidth. While the architecture targets future 6G networks, its feasibility is validated on a 5G Core-based prototype to emulate 6G in-network computing scenarios under realistic workloads. Our prototype evaluation demonstrates that the proposed architecture improves throughput, reduces latency and jitter, stabilizes data flow, and enhances AI-based video object detection performance.

Computing and networking platforms would become more decentralized toward B5G and 6G. Various functions and services will be deployed at the far edge, a site located physically closer to the user devices than the current edge-cloud design pattern in 5 G MEC. In such ultra-distributed environments, efficient use of those finite and limited computing and networking resources is one of the most important requirements for telecom operators. UPF, one component of 5GC, is a key function that links mobile and service domains. UPF is a typical example deployed in various platforms, from far edge to on-premise central data centers and clouds. DPUs are an advanced form of SmartNICs that are good at network, security, and storage communication processing and can be controlled as one of the in-network computing infrastructures. In this paper, we present the future telecom network architecture by offloading various network functions to DPUs and implementing UPF features as a first target. We also identify the feasibility of scalable UPF configurations from small to large scale leveraging DPUs and the DPU-dependent implementation issues for commercial applications by conducting essential functionality evaluation and carrier-grade large-scale performance tests.

Since cloud technology is rapidly advancing and spreading, is playing an increasingly important role in society, and will be widely used as the foundation for mobile network systems, innovative concepts called in-network computing (INC), in which computing functions are deployed, are widely considered for network and service functions. One of the benefits provided by INC is the optimization of computing and networking processing stacks. In this paper, we propose a system architecture for a mobile network equipped with computing capabilities called the Innetwork Service Acceleration Platform (ISAP). This enables end-to-end service performance acceleration, realizing processing optimization. We also present a prototype implementation of ISAP that integrates network and computing functions into a shared data plane on the basis of our proposed architecture, which includes a 3GPP standard-compliant 5G core network, server-less orchestration mechanisms, and multiple accelerator cards. Evaluations, including simulation and performance measurements, show ISAP has advantages over conventional designs, such as improved resource utilization efficiency, low latency, and deterministic service and network processing. These results validate our proposed scheme for realizing 6G INC.

Since cloud technology will permeate mobile networks in 6G, scenarios in which computing functions are deployed in the network are widely considered for not only network functions but also service functions. In this paper, we propose a system architecture, a mobile network equipped with computing capabilities, which enables end-to-end service performance acceleration. We evaluated our proposed architecture through evaluations and observed advantages over conventional designs, such as improved resource utilization efficiency and extremely low-latency service and network processing. The prototype implementation of integrating network and computing functions on the basis of our proposed architecture is also presented, which includes a 3GPP standard-compliant 5G core network, serverless orchestration mechanisms, and multiple accelerator cards. With the prototype above, we demonstrate how our proposed architecture federates network and service functions control and how it accelerates the application processing performance.

The orchestration and management of 5G network slicing (NS) requires cross-domain orchestration across 5G radio access network (RAN), 5G core network (CN), and transport network and also coordination of closed loops of each 5G network segment. We propose a resource orchestration and management architecture for end-to-end (E2E) 5G network slices to automate the flexible and high-performance NS’s management and orchestration. Through developing a prototype of the proposed architecture on an experimental 5G network, the effectiveness of our architecture is proven through prominent use cases certified by ETSI ZSM.

Network functions that apply advanced cloud-native technologies are called Cloud-native Network Functions (CNFs). CNFs reap many of the benefits of a microservices architecture. However, CNFs are expected to be used, for example, as a platform for MEC and will require more distributed deployment in various cloud environments from the edge to the private or public cloud than ordinary web services. As a result, the number of combinations of software and hardware resources will explode, making it difficult to design optimal hardware resources in accordance with the requirements of the various network services. To overcome this challenge, we propose an automated CNF provisioning engine that optimizes the hardware resources allocated to CNFs from the viewpoint of performance assurance and minimizing equipment costs even in various clouds. In this paper, we used machine learning for the cloud-native 5G mobile core to build a performance prediction model for both control plane and user plane functions under various hardware conditions on the basis of the performance characteristics data obtained from our test platform. From evaluating the prediction accuracy of the constructed models, we clarify that the models can predict with high accuracy even using features that can be easily fed back to the hardware resource design.

Cloud-native Network Functions (CNFs), which are an advanced form of Virtualized Network Functions (VNFs), are expected to optimize network slicing and make the development and deployment cycle of network services faster than VNFs by using a container-based microservices architecture. However, to satisfy the requirements for high reliability and performance assurance unique to network services, additional functions specific to the Kubernetes-based CNF platform are required. In this demonstration, we propose an automated CNF provisioning engine that optimizes the hardware resources of CNFs in multi-cloud environments on the basis of performance prediction and demonstrate its effectiveness by implementing it on Kubernetes as a core function of the future CNF platform.