Standards-Aligned 5G Platform Dynamically Shifts Computing Tasks Across Cloud and Edge

A research team from Ikerlan Technology Research Centre and the University of the Basque Country has developed and demonstrated FABRIC, an architectural orchestration framework that enables computational workloads to migrate dynamically between centralized cloud servers and localized edge nodes while preserving active mobile connectivity. Operating across distributed environments supported by fifth-generation and beyond-fifth-generation communication systems, the compute continuum requires coordinating processing resources across end-user equipment, intermediate access nodes, and central data repositories. In current telecommunication deployments, managing computational execution remains largely decoupled from core network routing controls, forcing operators to execute workload placement through proprietary tools or static configurations that fail to adjust when mobile execution contexts shift. By incorporating standardized interfaces established by the 3rd Generation Partnership Project and the European Telecommunications Standards Institute, the new architecture integrates dynamic user-plane routing directly with automated workload re-instantiation, allowing software components to migrate across physical boundaries without terminating live data streams.

The fundamental challenge confronting distributed compute continuum environments stems from the technical divide separating application lifecycle management from cellular network fabric reconfiguration. While modern container platforms like Kubernetes effectively orchestrate software execution across server clusters, they lack native mechanisms to instruct telecommunication cores to adjust data-plane tunnels toward newly assigned processing endpoints. Similarly, initiatives that publish network access points through application programming interfaces often operate independently across distinct telecommunication providers and cloud hosts. When latency-sensitive tasks follow mobile users toward regional edge servers, conventional fifth-generation network architectures demand manual or cumbersome operational steps to redirect the underlying traffic flows. Without unified coordination across computational demands, infrastructure performance indicators, and core network functions, distributed deployments face structural constraints in maintaining end-to-end quality of service, creating an operational bottleneck that hinders the transition from static hosting toward fully adaptive, responsive computing fabrics.

To overcome these limitations, the authors structured FABRIC around an architecture that couples computational workload re-instantiation with dynamic user plane function management across a three-tier execution hierarchy comprising cloud, edge, and user equipment layers. The platform coordinates four primary operational modules derived directly from telecommunications standards: an Application Service Provider, an Edge Computing Service Provider, a Public Land Mobile Network Operator entity, and an authoritative Network Resource Model. Under this framework, the Application Service Provider evaluates operational telemetry gathered from running workloads to determine whether tasks should migrate closer to end devices to reduce transit delay or move toward centralized servers to leverage scalable processing capacity. Once a migration decision is reached, the Edge Computing Service Provider identifies the optimal host within the designated destination zone, while the Public Land Mobile Network Operator reconfigures core routing elements by provisioning an updated User Plane Function instance and adjusting GPRS Tunneling Protocol data paths without disrupting underlying application states.

The software modularity of the framework is anchored in formal telecommunications specifications to guarantee practical interoperability with existing commercial infrastructure and prospective sixth-generation network paradigms. Core network management interfaces adhere to 3GPP Technical Specifications 23.501 and 23.502, utilizing a service-based architecture that integrates seamlessly with containerized network functions. To monitor and deploy service components, the platform relies on standardized Provisioning Management Services outlined in TS 28.532 and edge management definitions from TS 28.538 and TS 28.541. Rather than relying on custom proprietary messaging formats, the architecture leverages established reference points, such as the N3 interface connecting radio access base stations to user plane functions and the N9 interface bridging multiple user plane functions via GPRS Tunneling Protocol encapsulation. Furthermore, the platform adheres to ETSI Network Function Virtualization Management and Orchestration standards, operating directly alongside container infrastructure service clusters to decouple decision-making logic from physical hardware.

To validate the operational capabilities of the platform under controlled physical conditions, the researchers assembled a complete end-to-end telecommunications testbed running on dedicated hardware. The radio access network was constructed using an Amarisoft 5G New Radio stand-alone base station compliant with 3GPP Release 17 specifications, broadcasting across the n78 frequency band. Core network capabilities were powered by an extended instance of the open-source Free5GC software package running within standard container environments. The experimental computing continuum comprised two distinct kernel-based virtual machines hosted on separate physical machines linked via high-speed wired networking. The edge server was configured with 16 virtual processing cores and 32 gigabytes of random-access memory on dual Intel Xeon E5-2630 v4 processors, whereas the cloud server was provisioned with 32 virtual processing cores and 64 gigabytes of memory backed by an AMD EPYC 9554P processor, with both operating under Ubuntu 24.04.4 LTS.

The initial experimental validation focused on the physical resource behavior and migration latency of a synthetic compute-intensive workload generated through the stress-ng benchmarking utility. The benchmark was configured to run four central processing unit stress workers at full target utilization alongside one memory worker maintaining 4,096 megabytes of mapped space and two input-output processes over a 600-second execution window. The workload was initially started on the centralized cloud virtual machine and subsequently transferred to the edge virtual machine. System telemetry tracked central processing unit and memory load across both nodes, using a 12-second trailing moving average for processor calculations. The results revealed that during execution, the workload consumed eight equivalent virtual cores on both platforms, registering as 25 percent utilization on the 32-core cloud server and 50 percent utilization on the 16-core edge server, confirming identical resource consumption across disparate hardware layers.

Detailed timing analysis during the compute-intensive benchmark revealed distinct execution profiles between the initial workload transfer and subsequent re-instantiations. The initial migration required an overall duration of 18.983 seconds, a figure primarily driven by step five of the workflow, which took 10.310 seconds to configure and establish the user plane function redirection pathway across the core network. However, once this initial tunnel infrastructure was provisioned, subsequent workload migrations between the cloud and edge layers operated significantly faster, yielding a mean total duration of 8.768 seconds with a standard deviation of 2.117 seconds across five successive runs. During these later migrations, step three required an average of 5.951 seconds, step four executed in 0.005 seconds, and step five completed in an average of 2.305 seconds. A continuous processing overlap between instances ensured that the initial compute task remained active until the destination container became operational, preventing dropped processing cycles.

Resource tracking during the benchmark demonstrated marked differences in release dynamics between processor capacity and allocated memory across the participating nodes. While central processing unit consumption transferred cleanly—dropping to near zero on the origin node the moment the destination instance achieved operational stability—random-access memory utilization exhibited a gradual release curve attributable to internal operating system cleanup and memory reclamation processes. The cloud machine maintained approximately 8 gigabytes of memory usage during local execution and stabilized near 3.8 gigabytes following offloading, whereas the edge node rose from a baseline of 2.6 gigabytes to 6.8 gigabytes upon receiving the task. Network latency was simultaneously evaluated through round-trip time measurements between a mobile device and the running instances, recording a mean round-trip latency of 30.06 milliseconds with a standard deviation of 9.48 milliseconds across 453 samples at the cloud, compared to 31.04 milliseconds with a standard deviation of 6.29 milliseconds across 402 samples at the edge.

The second evaluation examined active traffic steering using a live data collection pipeline featuring an external Crosscall Z5 smartphone operating as user equipment. A custom generator running on the phone transmitted a random integer between 1 and 100 every second via REST application programming interface requests directed to a Python collector initially hosted on the cloud server. During active transmission, the platform initiated a service re-instantiation to move the collector application onto the edge server. Packet inspection across the N3, N6, and N9 network interfaces using Wireshark confirmed seamless data steering: user packets originated with a constant source address of 10.60.2.4 directed to destination 10.44.0.1, while the outer GPRS Tunneling Protocol transport header shifted its destination address from 192.168.9.4 to 192.168.19.157, successfully re-anchoring the live session to the local edge instance without interrupting packet delivery.

The traffic redirection experiment demonstrated complete preservation of application availability with zero data loss throughout the operational handover. The measured time interval separating the final request processed by the cloud instance from the first request handled by the edge node was exactly 0.578 seconds. Because this transition window remained substantially below the one-second transmission interval of the mobile client, all transmitted values arrived intact, achieving an absolute 100 percent request success rate. Total execution time for this single re-instantiation reached 23.746 seconds, including 13.123 seconds for preliminary initialization, 0.008 seconds for step four, 8.492 seconds for user plane path deployment, and 2.123 seconds for finalization. Application response times remained consistently between 70 and 98 milliseconds, while ping measurements showed a mean round-trip time of 31.81 milliseconds across 220 samples on the cloud and 30.91 milliseconds across 157 samples on the edge, confirming that dynamic re-instantiation introduces no measurable latency penalty.

While these experimental trials validate the fundamental mechanics of standards-aligned task re-instantiation and core user plane reconfiguration, the authors noted several operational constraints in the current evaluation scope. The empirical tests were conducted using manually triggered execution flows to isolate architectural behavior from algorithmic variance, omitting automated decision triggers such as telemetry anomaly detection or automated machine learning policies. Future iterations of the platform are slated to incorporate intelligent metaheuristic placement algorithms and real-time predictive telemetry capable of mitigating transient performance degradation before bottlenecks arise. Additionally, the researchers emphasize that industrial production deployments will require integrating robust security architectures, multi-vendor interoperability verifications, automated rollback protocols, and comprehensive compatibility assessments as sixth-generation mobile specifications mature under forthcoming 3GPP standards releases.

Subject of Research: Standards-aligned task orchestration and user-plane reconfiguration across the 5G compute continuum

Article Title: FABRIC: a standards-aligned orchestration platform for compute continuum operations in 5G networks and beyond

Article References: Nieto, G., Puerta-Beldarrain, M., Blanco, B., Aldecoa, E., Costa, I., Lopez-Novoa, U., & Liberal, F. (2026). FABRIC: a standards-aligned orchestration platform for compute continuum operations in 5G networks and beyond. Frontiers in Computer Science, 8, Article 1942828. https://doi.org/10.3389/fcomp.2026.1942828

Image Credits: AI Generated

DOI: 10.3389/fcomp.2026.1942828

Keywords: 5G networks, compute continuum, edge computing, cloud orchestration, 3GPP standards, ETSI NFV, user plane function, workload migration, GTP-U tunneling, service continuity, telecommunications, stand-alone 5G, packet data units, network virtualization, distributed computing, latency optimization