Robustness of Multiple High Speed TCP CUBIC Connections under Severe Operating Conditions

Abstract

We study the adaptation capabilities and robustness of the high-speed TCP CUBIC algorithm. For this purpose we consider a network environment with variable and high random packet loss and a large Bandwidth-Delay product, shared by multiple heterogeneous TCP connections. The analysis is based on and supported by packet-level simulations. The results show that the aggressive nature of CUBIC's nonlinear congestion window control principle causes a degradation of the time-average throughput at the moderate level of random packet loss even under increasing Round-Trip-Time of the flow. However, this algorithmic scalability and loss-free-time-dependent window growth allows recovering transmission rate faster in the high packet loss region due to the statistically lower number of dropped packets, compared to the moderate loss level.