Scalable Uplink Signal Detection in C-RANs via Randomized Gaussian Message Passing

Xiaojun Yuan, Congmin Fan, Xiaojun Yuan 0002, Ying Jun ZhangIEEE Trans. Wireless Communications 2017 · DOI

TL;DR

This paper tackles the challenge of computationally intensive signal detection in Cloud Radio Access Networks (C-RANs) as they scale to massive sizes. The authors introduce a novel Randomized Gaussian Message Passing (RGMP) algorithm that significantly improves convergence probability and maintains constant computational complexity per unit area, making C-RAN signal processing truly scalable.

What Problem Does This Paper Address?

Next-generation wireless networks are increasingly adopting the Cloud Radio Access Network (C-RAN) architecture. C-RANs centralize baseband processing in a data center, connecting it to numerous remote radio heads (RRHs) via high-bandwidth fiber. This architecture promises enhanced capacity and flexibility, especially through coordinated signal processing across RRHs. However, as the number of RRHs and users grows, the computational complexity of traditional signal detection methods (like MMSE) becomes prohibitively high, scaling cubically or quadratically with network size. This poses a significant barrier to realizing the full potential of massive C-RAN deployments. The paper aims to develop a signal detection algorithm that is "scalable," meaning its performance is near-optimal, its computational complexity per RRH remains constant as the network grows, and its total computation time (especially with parallelization) also remains constant.

Methods

The core of the paper lies in developing a scalable signal detection algorithm for C-RANs. The authors first model the C-RAN uplink signal detection problem as an inference problem over a bipartite random geometric graph, where RRHs and users are nodes and edges represent connections within a certain distance threshold (channel sparsification). This graph structure is then transformed into a factor graph.

The paper proposes a Randomized Gaussian Message Passing (RGMP) algorithm. Traditional Gaussian Message Passing (GMP) algorithms, while efficient on sparse graphs, struggle with convergence on the "loopy" and locally dense factor graphs generated by C-RANs. RGMP addresses this by introducing a randomized serial message updating schedule. Instead of updating messages simultaneously or in a fixed order, RGMP updates them in a random sequence. This randomization is shown to weaken the detrimental effects of loops in the graph, significantly improving convergence probability.

To enable parallel processing, RGMP is extended to a Blockwise RGMP (B-RGMP) algorithm. This algorithm divides message updates into multiple timeslots, allowing for parallel computation within each timeslot. The authors provide theoretical analysis for both RGMP and B-RGMP, deriving convergence conditions and demonstrating that B-RGMP offers improved convergence properties, especially for the case of two timeslots.

Datasets

Not discussed in the paper. The paper uses simulated data generated based on the described system model and channel characteristics.

Key Insights and Ideas

  • Randomization for Convergence: The key insight is that introducing randomness into the message update schedule of GMP can dramatically improve its convergence on loopy graphs, a common characteristic of C-RAN factor graphs. This is a novel approach to tackle a long-standing problem in message-passing algorithms.
  • Scalability via Sparsification and Randomization: By combining channel sparsification (reducing the number of connections) with randomized message passing, the authors achieve a truly scalable solution where computational complexity per unit area remains constant.
  • B-RGMP for Parallelism: The generalization to B-RGMP is crucial for practical implementation. It allows for parallel processing, ensuring that the total computation time doesn't increase with network size, a critical requirement for massive C-RANs.
  • Less Stringent Convergence Conditions: The paper analytically shows that RGMP and B-RGMP have less stringent convergence conditions compared to conventional GMP, meaning they are more likely to converge in real-world C-RAN scenarios.

Key Results

  • Improved Convergence Probability: RGMP converges with a much higher probability than conventional GMP, with simulations showing no divergence for RGMP in large networks.
  • Constant Iteration Count: The number of iterations required for RGMP to converge does not increase with network size, a key factor in achieving scalability.
  • Linear Total Complexity: The total computational complexity of RGMP is linear in the network size.
  • Constant Computation Time for B-RGMP: With parallel implementation and a fixed number of timeslots, B-RGMP achieves constant total computation time, regardless of network size.
  • Faster Convergence than Alternatives: RGMP and B-RGMP demonstrate significantly faster convergence speeds compared to algorithms like ADMM and GAMP, even when compared to damped GMP. For instance, RGMP achieves a relative error of 0.001 within 10 iterations for a network of 40 RRHs, while ADMM requires over 5000 iterations.

Limitations and Future Work

  • Damping Factor Optimization: For damped GMP, determining the optimal damping factor is still an open problem, and it often requires extensive simulations, which is impractical for large-scale C-RANs.
  • Randomized Message Updating Complexity: While effective, generating a large number of random variables for the update schedule in RGMP introduces some computational overhead. B-RGMP aims to mitigate this.
  • Future Research Directions: The authors suggest extending RGMP to scenarios with constellation constraints, limited fronthaul capacity, and for downlink beamforming in C-RANs. They also highlight the potential for improving message-passing algorithms by introducing randomized message updating in other contexts.

Summary updated

BibTeX
@article{yuan2017scalable,
  author = {Xiaojun Yuan and Congmin Fan and Xiaojun Yuan and Ying Jun Zhang},
  title = {{Scalable Uplink Signal Detection in C-RANs via Randomized Gaussian Message Passing}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2017},
  doi = {10.1109/twc.2017.2706680},
}
Figure