Abstract
This paper investigates the cross-layer challenges and performance of Hybrid Beamforming (HBF) and Multi-User Multiple-Input Multiple-Output (MU-MIMO) in 5G millimeter wave (mmWave) cellular networks. While traditional research focuses on link-level analysis under full-buffer conditions, this study evaluates these technologies within a full-stack TCP/IP environment. By extending the ns-3 mmWave module, the research reveals critical interplays between physical layer HBF techniques and higher-layer protocol behaviors, specifically regarding MAC scheduling and TCP congestion control. The results demonstrate that while Spatial Division Multiple Access (SDMA) can significantly increase throughput, achieving these gains requires navigating complex trade-offs between radio efficiency and synchronization.
Introduction
Millimeter-wave (mmWave) frequencies are the cornerstone of 5G’s high-capacity promise, offering massive bandwidth but suffering from significant pathloss. To overcome these propagation challenges, 5G systems utilize beamforming to focus energy into narrow, high-gain beams.
In practice, beamforming implementations fall into three categories:
- Analog Beamforming: Simple and cost-effective but limited to supporting a single beam/user at a time.
- Digital Beamforming: Highly flexible and powerful but prohibitively expensive and power-hungry for mmWave hardware.
- Hybrid Beamforming (HBF): A middle ground that combines analog phase shifters with digital baseband processing, enabling Multi-User MIMO (MU-MIMO) without the hardware overhead of fully digital systems.
Despite the theoretical potential of HBF, most existing literature evaluates its performance in isolation. This paper addresses the "full-stack gap" by analyzing how HBF interacts with real-world network protocols like TCP and MAC scheduling.
Methodology and System Design
To conduct this study, the authors extended the ns-3 mmWave module, creating an open-source full-stack simulator capable of modeling HBF MU-MIMO. The simulation environment follows 3GPP TR 38.901 standards, utilizing an Urban Macro (UMa) scenario at 28 GHz with realistic TCP/IP traffic.
HBF Design Strategies
The study evaluates two Minimum Mean Squared Error (MMSE) precoding designs:
- Frequency-Flat (FMBF): This approach calculates a single precoding matrix for the entire bandwidth. It minimizes feedback overhead but may sacrifice performance in highly dispersive channels.
- Frequency-Selective (SMBF): This design calculates unique precoding weights for different subcarriers. While it offers superior performance, it requires significantly higher feedback overhead from the User Equipment (UE) to the Base Station (BS).
MAC Layer Scheduling
A primary contribution of this work is the identification of a "scheduling conflict" in MU-MIMO systems. To manage this, two novel schedulers were proposed:
- Padded mmWave Round Robin (PMRS): This scheduler forces all users in a scheduled "bundle" to start their transmissions simultaneously. This synchronization allows the BS to calculate perfect interference-nulling weights, though it may waste resources on "padding" if users have unequal data volumes.
- Asynchronous mmWave almost-RR (AMRS): This scheduler prioritizes resource efficiency by allowing users to begin transmission as soon as they are ready. However, this creates a dynamic interference environment where the BS cannot easily "null" a user that joins mid-transmission.
Key Insights: The Cross-Layer Interplay
The Synchrony Requirement
The research highlights that for MMSE beamforming to effectively cancel interference, the BS requires stable channel state information for all active users at the start of a transmission interval. In asynchronous scenarios (AMRS), the introduction of a new user mid-stream causes sudden SINR drops for existing users, as the beams cannot be updated instantaneously to account for the new interference source.
Uplink Vulnerability
Interference management is significantly more challenging in the Uplink (UL) than the Downlink (DL). In the UL, the "near-far problem" is exacerbated; a user close to the BS can completely overwhelm the signal of a distant user if their beams are not perfectly orthogonal. This makes synchronous scheduling (PMRS) almost mandatory for viable UL MU-MIMO performance.
TCP Sensitivity
The study reveals that TCP is highly sensitive to the transient SINR outages caused by asynchronous scheduling. Even if the physical layer remains high-speed, TCP often interprets interference-related packet loss as network congestion, triggering back-off mechanisms that unnecessarily throttle the connection throughput.
Results and Performance Evaluation
- Capacity Gains: Transitioning from a traditional single-layer TDMA system to a 4-layer MU-MIMO HBF configuration can approximately double the total network throughput.
- Scheduler Efficacy: The synchronous PMRS scheduler provided the most stable performance, achieving throughputs exceeding 420 Mbps in DL and 450 Mbps in UL for 4-layer configurations.
- Asynchrony Failure in UL: While AMRS performed adequately in the Downlink, it struggled in the Uplink, where throughput plummeted to ~250 Mbps due to unmanaged inter-user interference.
- Latency: The system maintained low latency, with over 75% of packets delivered in under 20ms in the 4-layer configuration, meeting the requirements for modern real-time applications.
Conclusion and Future Directions
Hybrid Beamforming is essential for unlocking the capacity of 5G mmWave, but its success depends heavily on the coordination between the physical and MAC layers. While frequency-selective HBF (SMBF) and synchronous scheduling (PMRS) offer the best performance, they come with increased feedback overhead and potential resource waste.
Future research should explore mitigating these trade-offs through:
- Non-Orthogonal Multiple Access (NOMA): Integrating NOMA to further enhance spectral efficiency.
- Joint Decoding: Implementing advanced receivers (e.g., Sphere Decoding) at the BS to better handle asynchronous interference.
- Imperfect CSI: Evaluating performance under realistic channel estimation errors and feedback quantization.
Summary updated