Joint UPF and Edge Applications Placement and Routing in 5G & Beyond

TL;DR

The paper tackles the problem of jointly orchestrating UPF deployment, edge applications (EAs), and user-demand routing in 5G+ edge networks to maximize operator profit under latency and resource constraints. It presents an ILP formulation and proves NP-hardness, then introduces RanGr, a greedy ranking-based heuristic that anchors and places demands while routing through precomputed paths, striving to minimize active edge nodes. Through extensive simulations on small, medium, and large topologies, RanGr achieves near-optimal profit (about a 3% gap from the ILP) and outperforms Greedy and Top-K benchmarks in overloaded scenarios, while maintaining high CPU utilization and low runtimes. The work demonstrates a scalable, revenue-earning orchestration framework applicable to operators and campus/private networks, with future directions including energy-aware and dynamic-demand extensions.

Abstract

The development of 5G networks has enabled support for a vast number of applications with stringent traffic requirements, both in terms of communication and computation. Furthermore, the proximity of the entities, such as edge servers and User Plane Functions (UPFs) that provide these resources is of paramount importance. However, with the ever-increasing demand from these applications, operators often find their resources insufficient to accommodate all requests. Some of these demands can be forwarded to external entities, not owned by the operator. This introduces a cost, reducing the operator's profit. Hence, to maximize operator's profit, it is important to place the demands optimally in internal or external edge nodes. To this end, we formulate a constrained optimization problem that captures this objective and the inter-play between different parameters, which turns out to be NP-hard. Therefore, we resort to proposing a heuristic algorithm which ranks the demands according to their value to the operator and amount of resources they need. Results show that our approach outperforms the benchmark algorithms, deviating from the optimal solution by only ~3% on average.

Figures (3)

• Figure 1: System overview depicting the initialization of UPFs and EAs, and the placement of demands in the OENs and the EEN. As resources in OENs become scarce, demands are only anchored in the UPFs and are then offloaded to the EEN.
• Figure 2: Edge network topologies considered for the evaluation. The configuration of the number of BSs (triangles), switches (circles), and OENs (squares) for each topology is: a) $(4, 6, 2)$, b) $(8, 8, 4)$, and c) $(12, 10, 6)$. The same amount of resources is allocated to each OEN and link in each topology.
• Figure 3: Performance of the algorithms for different configurations of input demands, in terms of a) total profit, b) OEN CPU utilization, and c) number of demands placed on OENs. Results for the small, medium and large topologies are shown in dotted, dashed, and solid lines, respectively.

Theorems & Definitions (3)

• Lemma 1
• proof
• Definition 3