TCP ROCCET: An RTT-Oriented CUBIC Congestion Control Extension for 5G and Beyond Networks

TL;DR

TCP ROCCET addresses bufferbloat in loss-based TCP CUBIC within 5G and beyond by introducing an RTT-oriented extension that tracks $sRTT$ dynamics and ACK-rate, partitioning operation into LAUNCH and ORBITER phases. The approach uses $srRTT$ and $cum\_cwnd$ based indicators to detect congestion and apply controlled cwnd reductions, enabling faster adaptation to capacity changes while reducing latency. Real-world 5G measurements show ROCCET achieves latency reductions with comparable throughput to BBRv3, and bandwidth-sharing analyses demonstrate ROCCET’s deployability due to fairer behavior among ROCCET flows and less disruption when competing with other algorithms. The work provides a practical, kernel-extendable solution for 5G networks with configurable parameters and a roadmap toward official Linux integration.

Abstract

The behavior of loss-based TCP congestion control algorithms like TCP CUBIC continues to be a challenge in modern cellular networks. Due to the large RLC layer buffers required to deal with short-term changes in channel capacity, the behavior of both the Slow Start and congestion avoidance phases may be heavily impacted by the lack of packet losses and the resulting bufferbloat. While existing congestion control algorithms like TCP BBR do tend to perform better even in the presence of large bottleneck buffers, they still tend to fill the buffer more than necessary and can have fairness issues when compared to loss-based algorithms. In this paper, we analyze the issues with the use of loss-based congestion control algorithms by analyzing TCP CUBIC, which is currently the most popular variant. To mitigate the issues experienced by TCP CUBIC in cellular networks, we introduce TCP ROCCET, a latency-based extension of TCP CUBIC that responds to network congestion based on round-trip time in addition to packet loss. Our findings show that TCP ROCCET can reduce latency and bufferbloat compared to the standard CUBIC implementation, without requiring a specific network architecture. Compared to TCP BBRv3, ROCCET offers similar throughput while maintaining lower overall latency. The evaluation was conducted in real 5G networks, including both stationary and mobile scenarios, confirming ROCCET's improved response to network congestion under varying conditions.

Figures (9)

• Figure 1: Characteristic cwnd growth function of TCP CUBIC. With the cwnd size before the last congestion event ($W_{max}$) as a saddle point. The saddle point is reached at time $K$ after the last congestion event.
• Figure 2: TCP CUBIC first gets stuck in Slow Start and then in congestion avoidance during an unwanted side effect of the Linux kernel implementation. The two loss-based CE in the first seven seconds of the transmission are caused by L2 mechanisms and are not related to congestion.
• Figure 3: Schematic flow chart of the *ROCCET algorithm. Split in the Slow Start phase extension *LAUNCH and the congestion control phase *ORBITER. Gray states are end states.
• Figure 4: Measurement topology for mobile (standalone) and stationary downlink (non-standalone) measurements in two commercial NRnetworks.
• Figure 5: Goodput and sRTT boxplots for stationary 5G standalone greedy TCP throughput measurements in the downlink direction. Each measurement is 30 s long; during the run, the congestion control was alternated. We did 120 measurements for each congestion control.