NeuralEmu: in situ Measurement-Driven, ML-based, High-Fidelity 5G Network Emulation

1. Problem Statement

The evaluation of next-generation network algorithms (e.g., congestion control, adaptive bitrate streaming) in 5G environments is currently hindered by a "reality gap" between existing emulation tools and live commercial networks.

• Limitations of Record-and-Replay (e.g., Mahimahi, CellReplay): These tools capture a static trace of network behavior under a specific, often aggressive workload (saturator). They fail to capture the closed-loop feedback between the application and the 5G base station scheduler. In reality, the scheduler allocates resources dynamically based on real-time buffer occupancy and channel conditions. Record-and-replay tools break this causality, leading to inaccurate evaluations where applications behave differently than in a live network (e.g., overestimating available bandwidth).
• Limitations of Full-Stack Simulators (e.g., ns-3, srsRAN): These systems model the entire protocol stack but rely on simplified, open-source scheduling logics (e.g., Round-Robin, proportional fair) that cannot replicate the complex, proprietary, and vendor-specific scheduling algorithms of commercial 5G base stations. Additionally, they suffer from high computational overhead.

Core Challenge: The 5G scheduler is a "black box" that maps instantaneous user states (buffer occupancy, Channel Quality Indicator – CQI) to resource allocations (PRBs, MCS) in real time. Replicating this behavior without access to the proprietary scheduler code or the hidden internal states of background users is the primary hurdle.

2. Methodology: NeuralEmu Framework

NeuralEmu proposes a scheduler-driven digital twin approach. Instead of simulating the entire stack or replaying static traces, it uses Machine Learning (ML) to learn the mapping function of the commercial 5G scheduler directly from high-resolution, in situ measurements.

The system consists of three main components:

A. High-Resolution Data Collection

NeuralEmu uses a synchronized measurement framework to collect data at the application, device, and radio levels:

• Network View: Uses software-defined radios (NR-Scope) to passively capture downlink control information and record the scheduler's final decisions (PRB allocation and MCS) for all users in the cell.
• Device View: Uses tools like CellNinja on controlled, tethered smartphones to extract local CQI reports and modem protocols.
• Buffer Inference: Since the base station's internal buffer occupancy is opaque, NeuralEmu infers it by analyzing packet timings. It calculates the difference between the server egress time, the internet path delay (estimated via minimum one-way delay), and the modem ingress time to estimate the time packets waited in the RAN buffer.

B. Traffic Reconstructor (Solving the "Missing Data" Problem)

To train a scheduler model, the state of all users must be known, including uncontrolled background users whose traffic patterns are invisible.

• Inverse Problem: The Traffic Reconstructor is an ML model that works backward from observable scheduling outcomes (PRBs, MCS, Transport Block Size) to derive the hidden inputs: the input traffic patterns and CQI of background users.
• Mechanism: By analyzing the aggregated resource allocation and the known states of controlled users, the model derives the buffer dynamics and channel conditions of the "noise" users, effectively reconstructing the full cell state.

C. Neural Scheduler (The Emulation Engine)

This is the core ML model that acts as the emulator.

• Architecture: It is split into two specialized sub-models to handle the dual nature of cellular resource management:
  1. PRB Model: Uses a Gated Recurrent Unit (GRU) for temporal features and a Transformer encoder for cross-user attention. It predicts how limited Physical Resource Blocks (PRBs) should be split among competing users based on their relative buffer states and channel qualities.
  2. MCS Model: Uses a Transformer to predict the Modulation and Coding Scheme (MCS) for each user primarily based on their individual CQI.
• Operation: The model takes the reconstructed state (buffer occupancy, CQI) of all users and predicts resource allocation for the next time slots (granularity of 500 µs).

D. Implementation

NeuralEmu is implemented as a high-performance Linux middlebox using AF_XDP sockets and NVIDIA TensorRT for low-latency inference (<500 µs). It enforces TDD frame structures and simulates HARQ retransmissions to replicate delay jitter.

3. Key Contributions

1. First Multi-User ML Emulator: NeuralEmu is the first cellular emulator capable of accurately handling multiple simultaneous users sharing a base station and capturing the dynamic conflicts between users.
2. Scheduler-Driven Paradigm: It shifts the focus of emulation from the physical layer or static traces to the MAC scheduler logic, bridging the gap between theoretical design and commercial reality.
3. Traffic Reconstruction: Introduces a novel ML approach to invert scheduling telemetry and recover hidden background user states, solving the critical problem of data visibility.
4. Open Ecosystem: The authors commit to open-sourcing the software, high-resolution in situ telemetry datasets, and pre-trained models for future research.

4. Results and Evaluation

The authors evaluated NeuralEmu against live T-Mobile 5G networks and state-of-the-art tools (CellReplay, Mahimahi) across various applications:

• Congestion Control: NeuralEmu significantly reduced emulation errors for throughput and latency. For example, it reduced the RTT emulation error for the Copa algorithm by over 80% compared to CellReplay.
• Web Page Load Time: NeuralEmu reduced the emulation error by 55% compared to CellReplay and by 51% compared to Mahimahi.
• Video Streaming (VoD): In scenarios with background traffic (e.g., a BBR flow competing with a VoD client), NeuralEmu accurately replicated dynamic resource sharing, whereas record-and-replay tools failed to capture the scheduler's response to the "polite" nature of VoD traffic. NeuralEmu reduced ABR bitrate emulation errors from ~7% to <1%.
• WebRTC & Cloud Gaming: NeuralEmu achieved a 57% reduction in WebRTC encoder bitrate error and a 51% reduction in cloud gaming one-way latency error compared to existing tools.
• Performance: Inference time on a consumer GPU is approximately 456 µs (median), enabling it to keep pace with real-time 5G slot scheduling (0.5 ms).

5. Significance

NeuralEmu represents a paradigm shift in network emulation. By treating the 5G scheduler as a learnable function rather than a static environment or simplified algorithm, it provides a standardized, high-fidelity testbed for network protocols.

• Impact on Research: It enables researchers to rapidly iterate on congestion control and ABR algorithms without incurring the logistical nightmare of live field trials or the inaccuracies of simulations.
• Realism: It is the first tool to faithfully replicate the "closed-loop" interaction where application behavior influences network scheduling, which in turn affects application performance.
• Future-Proofing: As 5G evolves and vendors deploy new scheduling algorithms, the ML-based approach allows NeuralEmu to be retrained with new telemetry data to maintain relevance without needing to reverse-engineer proprietary code.