PaperlayerView exact revision on arXiv

Source-linked AI summary

Edge Computing Aware NOMA for 5G Networks

Abbas Kiani, Nirwan Ansari

arXiv:1712.04980v1cs.NI

TL;DR

The paper addresses the lack of NOMA-aware joint optimization for MEC communication and computing resources in 5G. It proposes an edge-computing-aware NOMA framework with clustering, RB allocation, and power control, and simulations report reduced energy consumption, near-optimal heuristic performance, and fairness effects from computing-RB granularity.

  • Problem

    The paper addresses the absence of prior work studying NOMA’s advantages for edge computing and the lack of joint NOMA optimization of communication and computing resources with intra-cell interference.

  • Method

    The proposed framework jointly optimizes NOMA user clustering, frequency and computing RB allocation, and transmit powers, using heuristic allocation and independent convex power control per cluster.

  • Results

    Simulations show that the heuristic has energy consumption close to the optimal approach, while its computation time remains within a few seconds as the optimal approach grows rapidly with users.

  • Takeaways & Limitations

    Increasing available frequency RBs or the maximum number of users per RB reduces energy consumption, while finer-granularity computing RBs produce fairer allocations.

Abstract

from arXiv · show

With the fast development of Internet of things (IoT), the fifth generation (5G) wireless networks need to provide massive connectivity of IoT devices and meet the demand for low latency. To satisfy these requirements, Non-Orthogonal Multiple Access (NOMA) has been recognized as a promising solution for 5G networks to significantly improve the network capacity. In parallel with the development of NOMA techniques, Mobile Edge Computing (MEC) is becoming one of the key emerging technologies to reduce the latency and improve the Quality of Service (QoS) for 5G networks. In order to capture the potential gains of NOMA in the context of MEC, this paper proposes an edge computing aware NOMA technique which can enjoy the benefits of uplink NOMA in reducing MEC users' uplink energy consumption. To this end, we formulate a NOMA based optimization framework which minimizes the energy consumption of MEC users via optimizing the user clustering, computing and communication resource allocation, and transmit powers. In particular, similar to frequency Resource Blocks (RBs), we divide the computing capacity available at the cloudlet to computing RBs. Accordingly, we explore the joint allocation of the frequency and computing RBs to the users that are assigned to different order indices within the NOMA clusters. We also design an efficient heuristic algorithm for user clustering and RBs allocation, and formulate a convex optimization problem for the power control to be solved independently per NOMA cluster. The performance of the proposed NOMA scheme is evaluated via simulations.

I. INTRODUCTION

The paper addresses the emerging combination of NOMA and MEC in 5G, targeting massive connectivity, low latency, and reduced uplink energy consumption. It proposes joint optimization of clustering and communication-computing resources, supported by heuristic allocation and cluster-wise power control.

  • Motivation: NOMA and MEC are proposed together to address 5G capacity, connectivity, latency, and edge-computing demands.NOMA is associated with spectral efficiency and capacity, while MEC supports low-latency computation offloading.
  • Research gap: Existing work had not jointly optimized computing and communication resources for NOMA while accounting for intra-cell interference.The paper positions this joint optimization as the gap it addresses.
  • Contributions: The proposed framework minimizes MEC users’ energy consumption by optimizing clustering, computing and communication resource allocation, and transmit powers.It introduces computing RBs alongside frequency RBs and considers task deadlines when establishing minimum data-rate requirements.
  • Contributions: The authors design a heuristic algorithm for user clustering and RB allocation and a convex power-control problem solved independently for each NOMA cluster.These methods separate resource allocation from cluster-level transmission power control.
  • Evaluation: The proposed scheme and heuristic algorithm are evaluated through extensive simulations to assess NOMA’s effect on MEC users’ uplink energy consumption.The stated evaluation focuses on the energy-saving benefits of uplink NOMA.

II. SYSTEM MODEL AND PROBLEM FORMULATION

The system model considers a single cell in which an eNB equipped with a cloudlet serves uniformly distributed edge-computing users, each associated with an offloaded task. Each task is characterized by its workload and input size.

  • System scenario: The model uses a single-cell scenario with one cloudlet-equipped eNB serving uniformly distributed edge-computing users.The users offload their tasks to the cloudlet through the eNB.
  • Task model: Each user task is represented by a workload λ_u measured in required CPU cycles and an input L_u measured in bits transferred to the eNB.Users and tasks are used interchangeably in the formulation.

A. Communication Resources

The communication model divides bandwidth into frequency RBs that NOMA users may share, groups users into ordered clusters, and models successive-interference cancellation and uplink transmission energy.

  • Frequency RBs: Available bandwidth is divided into frequency RBs R_f={1,...,M_f}, each with bandwidth B, and multiple users may share an RB under NOMA.RB allocation to clusters is represented by the binary variable β_{r,i}.
  • NOMA clustering: Users are assigned to NOMA clusters, with each cluster receiving a set of frequency RBs and at least two users sharing those RBs.The number of clusters is bounded by the number of users and the NOMA sharing requirement.
  • Decoding order: Successive interference cancellation orders users within each cluster, so the j-th decoded user experiences interference from users with higher order indices.The first decoded user sees interference from all later users in the cluster.
  • Rate and assignment model: The model uses binary assignment variables α_{i,j}^u to indicate users’ order positions, along with transmit powers, channel gains, and noise power in the achievable-rate formulation.Channel conditions vary across both users and frequency RBs.
  • Energy model: Uplink transmission time is determined by each task’s input size and achievable rate, and user energy consumption is then obtained from transmission power and time.The supplied formulation text introduces these quantities sequentially without displaying the complete equations.

B. Computing Resources

The computing model partitions cloudlet capacity into computing RBs and allocates them to order indices within NOMA clusters. Each computing RB represents a processing unit with capacity C CPU cycles per second.

  • Computing RBs: Cloudlet computing capacity is divided into computing RBs, analogous to communication resource blocks.A computing RB may represent a virtual machine or a CPU core.
  • Capacity allocation: Each computing RB provides C CPU cycles per second, and multiple computing RBs are allocated to each order index of each NOMA cluster.The number of RBs assigned to order index j in cluster i is denoted x_{i,j}.

III. OPTIMIZATION PROBLEM

The paper formulates an energy-minimization problem for MEC users under transmission and computing-time deadlines. The MINLP jointly represents clustering, resource allocation, ordering, and power-related decisions, motivating a two-phase solution.

  • The optimization minimizes total user energy consumption subject to transmission and computing-time constraints, including per-task deadlines.
  • Constraint C1 imposes a computing-aware minimum data-rate requirement, while C2 upper-bounds each task’s completion time by its deadline.
  • Constraints C3–C6 enforce unique cluster and order assignments, at least two users per cluster, exclusive frequency-RB allocation, and bounded computing-RB allocation.
  • Constraints C7–C12 enforce order-priority assignment, transmit-power budgets, positive powers, binary allocation variables, and integer computing-RB variables.
  • Because P1 is an intractable, generally non-convex MINLP, the authors use a two-phase approach beginning with heuristic clustering and RB allocation.

A. User Clustering and RBs allocation

Algorithm 1 separates user clustering, computing-RB allocation, and frequency-RB allocation. It prioritizes channel conditions and deadline-aware resource needs while accommodating an optional RB-contiguity extension.

  • User Clustering and RBs allocation: Algorithm 1 performs user clustering, computing-RB allocation, and frequency-RB allocation in three separated phases.
  • User Clustering and RBs allocation: Users are sorted by average channel condition, with higher-gain users assigned lower NOMA order indices to support SIC-based interference cancellation.
  • User Clustering and RBs allocation: Computing RBs are first allocated to satisfy Qu < Du, then remaining RBs are assigned to clusters using an objective-based favorite-cluster search.
  • User Clustering and RBs allocation: The computing-allocation comparison accounts for users’ input sizes and deadlines when selecting the favorite user.
  • User Clustering and RBs allocation: Frequency RBs initially satisfy users’ minimum data-rate requirements, after which remaining RBs are assigned by comparing candidate rate improvements across clusters.
  • User Clustering and RBs allocation: Under RB contiguity, frequency RB allocation becomes NP-hard, so the authors propose using RB groups as the allocation granularity.

B. Power Control

After clustering and RB allocation, the paper fixes assignment variables and optimizes transmit powers independently for each NOMA cluster. A variable transformation yields an equivalent problem that is convex in the high-SINR regime.

  • Power Control: With clustering and frequency-RB assignments fixed, integer constraints disappear and power control is formulated independently for each NOMA cluster.
  • Power Control: The power-control objective minimizes total power instead of energy because communication time is bounded by the computing-aware rate constraint.
  • Power Control: Problem P2 is equivalent to P3 after a one-to-one change of variables and introduction of auxiliary variables.
  • Power Control: The heuristic’s energy consumption is close to the optimum, while its solution takes a few seconds and the optimal method’s computation time grows rapidly with user count.
  • Power Control: Problem P3 is convex in the high-SINR regime, using the approximation log(1 + x) ≈ log(x) for the non-convex throughput constraint.
  • Power Control: The resulting power-control problem can be solved with efficient methods such as interior-point techniques, although complexity increases with cluster size.

IV. SIMULATION RESULTS

Simulations evaluate energy consumption, spectral efficiency, computing-time allocation, and fairness under varied communication and computing-resource configurations. The proposed scheme’s energy use generally benefits from more frequency resources, greater NOMA sharing, and finer computing-resource granularity.

  • Heuristic and optimal approaches: The heuristic algorithm achieves total energy consumption close to the optimal approach while producing a suboptimal solution within a few seconds.The optimal approach’s computation time grows rapidly with the number of users.
  • Frequency RBs and NOMA sharing: Energy consumption decreases as available frequency RBs increase and improves when umax rises from 1 to 2 or 3.The improvement is attributed to higher spectral efficiency, while larger intra-cell interference reduces the gap between higher umax values.
  • Spectral efficiency: Spectral efficiency increases with both the number of users and the average input of each user.Figure 3 evaluates the proposed scheme with umax = 3.
  • Computing-resource division: For higher user counts, energy consumption improves when computing capacity is divided into smaller RBs, using scenarios with 120 or 90 computing RBs.The paper attributes this to fairer allocation for users with different workloads.
  • Computing-time fairness: Computing RBs with 1 Giga cycles per second provide higher Jain’s computing-time fairness than RBs with 3 Giga cycles per second.The fairness index is bounded between 0 and 1 for the evaluated scenarios.

V. CONCLUSION

The paper proposes an edge computing aware NOMA technique for reducing MEC users’ energy consumption. It combines joint frequency and computing RB allocation with heuristic clustering and per-cluster convex power control, and evaluates the scheme through simulations.

  • The proposed technique leverages uplink NOMA to reduce MEC users’ energy consumption.
  • The optimization framework jointly optimizes user clustering, computing and communication resource allocation, and transmit powers.
  • The scheme allocates frequency and computing RBs to users assigned different order indices within NOMA clusters.
  • A heuristic algorithm handles user clustering and RB allocation, while convex power control is solved independently for each NOMA cluster.
  • Simulations evaluate and demonstrate the proposed NOMA scheme’s effectiveness in lowering energy consumption.
Loading 1712.04980v1…