CRRM: A 5G system-level simulator

TL;DR

CRRM introduces a Python-based 5G system-level simulator designed to close the usability gap for AI researchers. It replaces traditional discrete-event schedulers with a compute-on-demand graph of dependent blocks, achieving about a $2\times$ speed-up in typical mobility scenarios while preserving correctness. The tool integrates validated 3GPP TR 38.901 propagation models, 3-sector antenna patterns, subband interference coordination, and a tunable resource-allocation fairness parameter, with validation against PPP theory. This combination offers a practical, AI-friendly platform for rapid prototyping and policy evaluation in next-generation wireless networks.

Abstract

System-level simulation is indispensable for developing and testing novel algorithms for 5G and future wireless networks, yet a gap persists between the needs of the machine learning re- search community and the available tooling. To address this, we introduce the Cellular Radio Reference Model (CRRM), an open-source, pure Python simulator we designed specifically for speed, usability, and direct integration with modern AI frameworks. The core scientific contribution of CRRM lies in its architecture, which departs from traditional discrete-event simulation. We model the system as a set of inter-dependent computational blocks which form nodes in a directed graph. This enables a compute-on-demand mechanism we term smart update.

Figures (5)

• Figure 1: The CRRM computational data flow, with the shapes of the arrays indicated. Data propagates from root nodes like $U$ (representing UE locations), $C$, and $P$ (power), through a series of dependent blocks, each representing a distinct mathematical operation. The stacked blocks for RSRP, SINR, etc. illustrate the handling of multiple subbands. The horizontal red stripe indicates the rows of all the arrays requiring to be updated if a UE moves. Python advanced indexing is used to perform multiple UE move updates in one operation.
• Figure 2: Throughput as a function of distance for the RMa, UMa, UMi, and power-law pathloss models. The simulation captures the distinct decay characteristics of each propagation environment.
• Figure 3: Comparison of UE throughput for 1-sector (omnidirectional) and 3-sector antenna configurations. The plot shows the impact of the 3-sector antenna gain pattern as a UE moves in a circle around the base station.
• Figure 4: UE throughput as a function of the resource allocation fairness parameter, $p$. The plot illustrates how throughput is redistributed from high-spectral-efficiency users to low-spectral-efficiency users as $p$ increases.
• Figure 5: Validation of the simulated SIR distribution. The complementary cumulative distribution from a 1000-UE PPP simulation (dots) shows excellent agreement with the exact analytical theory (solid line).