A Low-Complexity Architecture for Multi-access Coded Caching Systems with Arbitrary User-cache Access Topology
Ting Yang, Minquan Cheng, Xinping Yi, Robert Caiming Qiu, Giuseppe Caire
TL;DR
This work tackles multi-access coded caching (MACC) under arbitrary user-cache access topologies by casting the delivery problem as a graph coloring task on a conflict graph derived from the MACC-PDA framework. It introduces a universal graph-based MACC formulation, shows that DSatur yields near-optimal delivery performance but is computationally heavy, and proposes a scalable GNN-based color predictor with a light postprocessing step to produce feasible, low-colorings. It also extends the index-coding (IC) converse bound to MACC with uncoded placement and proposes a low-complexity greedy bound that closely tracks the IC limit while dramatically reducing computation. Empirical results demonstrate that the GNN-based approach achieves delivery loads close to DSatur and the IC bound with substantial runtime savings, and that the greedy bound provides a practical, accurate approximation across large, irregular topologies. Overall, the framework enables scalable, near-optimal MACC delivery for diverse topologies, with strong potential for real-world edge caching deployments and dynamic networks.
Abstract
This paper studies the multi-access coded caching (MACC) problem under arbitrary user-cache access topologies, extending existing models that rely on highly structured and combinatorially designed connectivity. We consider a MACC system consisting of a single server, multiple cache nodes, and multiple user nodes. Each user can access an arbitrary subset of cache nodes to retrieve cached content. The objective is to design a general and low-complexity delivery scheme under fixed cache placement for arbitrary access topologies. We propose a universal graph-based framework for modeling the MACC delivery problem, where decoding conflicts among requested packets are captured by a conflict graph and the delivery design is reduced to a graph coloring problem. In this formulation, a lower transmission load corresponds to using fewer colors. The classical greedy coloring algorithm DSatur achieves a transmission load close to the index-coding converse bound, providing a tight benchmark, but its computational complexity becomes prohibitive for large-scale graphs. To overcome this limitation, we develop a learning-based framework using graph neural networks that efficiently constructs near-optimal coded multicast transmissions and generalizes across diverse access topologies and varying numbers of users. In addition, we extend the index-coding converse bound for uncoded cache placement to arbitrary access topologies and propose a low-complexity greedy approximation. Numerical results demonstrate that the proposed learning-based scheme achieves transmission loads close to those of DSatur and the converse bound while significantly reducing computational time.