Joint Radio–Computation Allocation

• Joint radio–computation resource allocation is a coordinated method that jointly assigns wireless (e.g., subcarriers, power) and computational (e.g., CPU cycles, VNF placement) resources to meet strict service-level objectives.
• It integrates diverse system models including MEC OFDMA, Tactile Internet, and CF-MIMO, leveraging alternating optimization, convex relaxation, and reinforcement learning for improved next-generation network performance.
• This joint approach significantly reduces energy consumption and latency while enhancing throughput and scalability compared to decoupled resource management methods.

Joint radio–computation resource allocation refers to the coordinated assignment of wireless communication resources (e.g., spectrum, transmit power, subcarriers, radio clusters) and computational resources (e.g., CPU cycles, computation time, placement of network functions, processing chains) to users, tasks, or network slices. This paradigm is fundamental to multi-tier edge computing, cloud-radio access networks, next-generation Tactile Internet, and cell-free massive MIMO networks, where both bandwidth and processing capacity are scarce and tightly coupled. By jointly optimizing these domains, systems achieve stringent service-level objectives (e.g., ultra-low latency, high throughput) more efficiently than possible with decoupled approaches.

1. Theoretical Foundations and System Models

Joint radio–computation resource allocation arises in network architectures where computation offloading or network function virtualization is performed in tandem with dynamic radio resource management.

Common canonical models include:

• Heterogeneous cellular/MEC network models: Macro base station (MBS) and multiple small base stations (SBS), each with local compute servers, interconnected via backhaul links, supporting users requesting end-to-end network services (service chains of VNFs) subject to subcarrier, transmit power, and NFV constraints (Gholipoor et al., 2019).
• OFDMA-based MEC systems: Single cloudlet or edge server with a set of subcarriers and limited CPU processing, servicing multiple single-task users who may select between local execution and task offloading, with energy and latency as critical constraints (Yu et al., 2016).
• RAN slicing: Edge servers host multiple logical slices (eMBB, URLLC, etc.), with resource blocks and CPU pools dynamically partitioned according to slice demand and policy (Zhou et al., 2021).
• Cloud-RAN and CF-MIMO: Remote radio heads (RRHs) or radio units (RUs) and a virtual base station (VBS) pool or distributed cluster processors (CPs), with jointly optimized radio clustering, beamforming, fronthaul routing, and processor placement (Tran et al., 2015, Li et al., 2023).

The joint allocation generally operates over highly coupled resource domains and discrete decisions (e.g., subcarrier/user mapping, offloading, VNF placement), yielding optimization problems that are inherently mixed-integer and nonconvex.

2. Joint Optimization Problem Formulations

At the core of these systems is an optimization problem that minimizes cost or maximizes utility subject to service constraints, where cost incorporates both radio and compute domains.

Typical formulations are:

System Type	Objective Function	Constraints (Selection)
MEC OFDMA Cloudlet	$$\min \sum_i (1-\alpha_i)E_i^{loc} + \alpha_i E_i^{tx}$$	- Subcarrier exclusivity <br> - CPU time budget <br> - Energy/deadline
Tactile Internet	$$\min C = \rho_1 (\sum p) + \rho_2 (\sum \text{NFV time})$$	- E2E delay bound <br> - Radio/compute assignment <br> - VNF scheduling
C-RAN w/ CPU Pool	$\max \sum_u q_u R_u(S,W)$	- Beam power <br> - RRH clustering <br> - VBS CPU capacity
RAN Slicing	$$\max w^{embb}b^{embb,avg} + w^{urllc}(d^{tar}-d^{urllc,avg})$$	- Resource block partition <br> - CPU allocation <br> - Slice isolation
CF-MIMO Fronthaul	$\min \max(\text{UL load, DL load})$	- Quantization scheduling <br> - Routing <br> - DU CPU/placement

Variables commonly include integer scheduling assignments (e.g., subcarrier or RB indices), continuous power or CPU allocations, offloading indicators, and potentially permutation variables for task or VNF execution order. Strict latency, cost, and resource-availability constraints make the problems NP-hard (Tran et al., 2015, Gholipoor et al., 2019, Yu et al., 2016, Li et al., 2023).

3. Algorithmic Approaches and Decomposition Techniques

Due to mixed-integer nonconvexity, all frameworks employ decomposition and iterative optimization.

Alternating/Block Coordinate Methods: Problems are decomposed into sequential subproblems conditioned on remaining variables, iterating until convergence:

• Subcarrier and transmit power are allocated given (fixed) compute assignments.
• CPU (or VNF) placement and scheduling are solved for fixed radio allocation.
• Joint subcarrier and CPU scheduling are integrated via greedy heuristics or dynamic programming to fully utilize resources (Yu et al., 2016).

Convex Relaxation and Successive Approximation:

• Nonconvex rate or resource constraints are handled with difference-of-convex (DC) programming, successive convex approximation (SCA), or second-order cone programming (SOCP) relaxations (Gholipoor et al., 2019, Tran et al., 2015, Li et al., 2022).
• Integer variables are temporarily relaxed, with explicit rounding/restoration at each main iteration (Li et al., 2022).

Matching and Game-Theoretic Allocation: For highly dynamic user-to-resource mapping (as in mmWave train-ground networks), two-sided matching games, swap-stability, and greedy admission heuristics are used for subchannel and compute assignment under energy constraints (Li et al., 2022).

Reinforcement Learning: In dynamic environments with partial knowledge, Markov decision process (MDP) models and (deep) Q-learning are used for adaptive joint allocation, either from scratch or with expert knowledge transfer to accelerate convergence and performance (Zhou et al., 2021, Dai et al., 2020). Knowledge transfer mechanisms warm-start the learning agent’s Q-table with expert radio allocation, requiring learning only the computation dimension in the joint allocation.

MILP Formulation with Modern Solvers: In large-scale CF-MIMO, the joint routing of fronthaul flows (UL/DL), quantization scheduling, and cluster processor assignment is captured as a pure Mixed Integer Linear Program and solved to global optimality with commercial solvers, leveraging modern computational power (Li et al., 2023).

4. Performance Benefits and Quantitative Results

A consistent conclusion across the literature is that joint resource allocation substantially outperforms decoupled or per-domain methods in the following key aspects:

• Cost and Energy: Joint allocation reduces total power consumption and computational cost. For Tactile Internet, total network cost in the 1 ms delay regime can be 30–50% lower than decoupled allocation, with the penalty for separate radio/NFV-RA growing as the E2E delay target tightens (Gholipoor et al., 2019). In MEC, coordinated joint allocation doubles the number of accepted offloading requests and energy savings versus per-resource optimization (Yu et al., 2016).
• Latency and SLA Adherence: Joint methods ensure end-to-end delay compliance (including both radio transmission, queueing, and VNF/computation times), a prerequisite in critical services such as URLLC and Tactile Internet (Gholipoor et al., 2019, Zhou et al., 2021). For RAN slicing, joint radio–computation allocation via knowledge transfer achieves up to 18.4% lower URLLC latency and 30.1% higher eMBB throughput than vanilla Q-learning (Zhou et al., 2021).
• Scalability and Utilization: In dynamic environments with varying user load, subcarrier pool, and bandwidth, algorithms fully utilize both bandwidth and computation pools, minimizing bottlenecks. For CF-MIMO, careful selection of UL quantization bits and processor placement allows halving fronthaul load with negligible physical layer spectral efficiency loss (Li et al., 2023).

5. Design Insights, Limitations, and Guidelines

Practical design insights that emerge from these studies include:

• Trade-off curves: Ultra-low-latency requirements invariably demand joint allocation. For example, for $D^{\max}<2$ ms in tactile applications, joint R-RA and NFV-RA is essential (savings >30%); for relaxed settings ($D^{\max}>5$ ms), separate optimization is sufficiently close to optimal (Gholipoor et al., 2019).
• Resource Pool Sizing: Over-provisioning of one domain (e.g., CPU at MEC) yields rapidly diminishing returns if the bottleneck resides in the other (radio), and vice versa. MEC server sizing should target the ratio of average traffic × processing coefficient to delay budget (Gholipoor et al., 2019).
• Impact of System Parameters:
  • Increasing subcarrier/RB pool increases multiuser diversity and reduces cost/noncompliance probability (Gholipoor et al., 2019, Zhou et al., 2021).
  • Tightening delay targets or decreasing quantization distortion (in UL) increases cost and reduces system capacity, but beyond a certain fidelity, extra radio/compute resources bring marginal return—optimal design operates at modest bit-depths or CPU allocations (Li et al., 2023).
  • For multi-tier or multi-agent systems, knowledge transfer–based RL dramatically accelerates resource adaptation under traffic and topology changes (Zhou et al., 2021).
• Algorithmic Efficiency and Convergence: Most alternating or block-coordinate schemes converge in tens of iterations, and the complexity per iteration is well characterized (e.g., $O((KN)^{3.5})$ for convex QP in (Li et al., 2022)).
• Structural Limitations: Existing models are often static (single-shot, ignore task arrivals), single-cell, and rely on perfect knowledge for deterministic optimization. Promising extensions include online dynamic arrivals, distributed multi-cell cooperation, and joint consideration of system energy and server operational costs (Yu et al., 2016).

6. Application Scenarios and Emerging Directions

Joint radio–computation resource allocation underpins several advanced wireless/networking use cases:

• Tactile Internet with E2E 1 ms latency (teleoperation, real-time control) (Gholipoor et al., 2019).
• 5G RAN slicing for eMBB and URLLC, dynamically adjusting to mixed traffic and slice-specific latency/throughput targets (Zhou et al., 2021).
• Mobile edge computing in dense/deployment-limited scenarios (small cloudlets, train–ground mmWave links, multi-tier networks with reconfigurable metasurfaces) (Li et al., 2022, Li et al., 2022).
• Cell-free massive MIMO deployments with constraint-aware fronthaul routing and distributed processor allocation (Li et al., 2023).
• Cloud-RAN architectures for cooperative beamforming and high system utility at constrained compute budgets (Tran et al., 2015).

Generalizations to deep transfer RL, multi-agent frameworks, multi-slice/multi-service multiplexing, and meta-surface-aided multi-tier offloading are active research directions with algorithmic and architectural implications.

7. Comparative Overview of Model Features

Reference	Application Domain	Radio Resource	Computation Resource	Optimization Approach	Key Result
(Gholipoor et al., 2019)	Tactile Internet, NFV	OFDMA UL/DL SCA	VNF chains, placement	ASM, SCA, heuristics	30–50% cost savings, E2E 1ms compliance
(Zhou et al., 2021)	5G RAN Slicing	RB allocation	MEC CPU fractions	RL/Transfer learning	18% latency, 30% throughput improvement
(Yu et al., 2016)	MEC Cloudlet, OFDMA	Subcarrier selection	Nonpreemptive CPU schedule	DP, greedy group allocation	2× offloading, energy vs decoupled
(Tran et al., 2015)	Cloud-RAN, C-RAN	Beamforming clusters	Central CPU pool	MI-SOCP, $\ell_1$ reweighting	>30% system utility, 2.5× rate over dist.
(Li et al., 2023)	CF-MIMO, Fronthaul	RU–UE clusters, quant	DU CPU placement, fronthaul	Global MILP	20% lower load, optimal quantization bits
(Li et al., 2022)	Multi-tier OFDMA/RMS	Subcarrier, power	Bit/time/task allocation	BCD, SCA, DC	30–50% energy saving over partial

All frameworks demonstrate that, in tightly coupled radio–compute environments, joint allocation strategies are indispensable for cost, energy, and performance optimization under modern application QoS guarantees.