This paper presents an analytical performance model for the TCP Congestion Avoidance algorithm, which predicts the bandwidth of a sustained TCP connection under light to moderate packet loss conditions. The model assumes that TCP avoids retransmission timeouts and has sufficient receiver window and sender data. It is validated through simulation and live Internet measurements, showing strong agreement with prior work in this area. The model is applied to bandwidth allocation in the Internet and demonstrates implications for the behavior of the Internet under high load from diverse user communities. The model is derived based on the stationary distribution of the congestion window of ideal TCP Congestion Avoidance under independent congestion signals with constant probability. A simplified derivation is presented, assuming periodic congestion signal losses, leading to the same mathematical form as the full derivation, though with a slightly different constant of proportionality. The model is applicable whenever TCP's performance is determined solely by the Congestion Avoidance algorithm. It is hypothesized to apply to nearly all implementations of SACK TCP under normal Internet conditions and to Reno TCP under more restrictive conditions. The model is tested through simulations of various TCP implementations under different loss conditions and queuing environments, including drop-tail and RED. It is also compared to live Internet measurements using diagnostic tools and real TCP implementations. The model is found to fit well in many scenarios, though it does not apply in situations where TCP performance is influenced by factors such as receiver window size, sender data availability, timeout behavior, or go-back-N behaviors. The model is used to predict TCP performance in different network conditions, including those with multiple congested gateways. It is shown to accurately predict bandwidth allocation in such scenarios. The model is also used to analyze the impact of packet loss on TCP performance, showing that high loss rates can significantly reduce throughput. The model is applied to real-world Internet measurements, showing that it can predict TCP behavior under various conditions. The model is found to be useful for predicting how TCP shares Internet bandwidth and its effects on the Internet. It represents an equilibrium process between loss, delay, and bandwidth. The model is most accurate when using delay and loss instruments in the TCP itself or when loss is randomized at the bottleneck. It is less accurate in cases with non-randomized losses, such as drop-tail queues. The model is also used to analyze the impact of different TCP implementations on performance, showing that FACK-RH TCP fits the model across a wide range of conditions, while Reno TCP deviates from the model under ordinary conditions. The paper concludes that the model can be used to predict the bandwidth of Congestion Avoidance-based TCP implementations under many conditions, and that it provides valuable insights into the behavior of the Internet under high load from diverse user communities.This paper presents an analytical performance model for the TCP Congestion Avoidance algorithm, which predicts the bandwidth of a sustained TCP connection under light to moderate packet loss conditions. The model assumes that TCP avoids retransmission timeouts and has sufficient receiver window and sender data. It is validated through simulation and live Internet measurements, showing strong agreement with prior work in this area. The model is applied to bandwidth allocation in the Internet and demonstrates implications for the behavior of the Internet under high load from diverse user communities. The model is derived based on the stationary distribution of the congestion window of ideal TCP Congestion Avoidance under independent congestion signals with constant probability. A simplified derivation is presented, assuming periodic congestion signal losses, leading to the same mathematical form as the full derivation, though with a slightly different constant of proportionality. The model is applicable whenever TCP's performance is determined solely by the Congestion Avoidance algorithm. It is hypothesized to apply to nearly all implementations of SACK TCP under normal Internet conditions and to Reno TCP under more restrictive conditions. The model is tested through simulations of various TCP implementations under different loss conditions and queuing environments, including drop-tail and RED. It is also compared to live Internet measurements using diagnostic tools and real TCP implementations. The model is found to fit well in many scenarios, though it does not apply in situations where TCP performance is influenced by factors such as receiver window size, sender data availability, timeout behavior, or go-back-N behaviors. The model is used to predict TCP performance in different network conditions, including those with multiple congested gateways. It is shown to accurately predict bandwidth allocation in such scenarios. The model is also used to analyze the impact of packet loss on TCP performance, showing that high loss rates can significantly reduce throughput. The model is applied to real-world Internet measurements, showing that it can predict TCP behavior under various conditions. The model is found to be useful for predicting how TCP shares Internet bandwidth and its effects on the Internet. It represents an equilibrium process between loss, delay, and bandwidth. The model is most accurate when using delay and loss instruments in the TCP itself or when loss is randomized at the bottleneck. It is less accurate in cases with non-randomized losses, such as drop-tail queues. The model is also used to analyze the impact of different TCP implementations on performance, showing that FACK-RH TCP fits the model across a wide range of conditions, while Reno TCP deviates from the model under ordinary conditions. The paper concludes that the model can be used to predict the bandwidth of Congestion Avoidance-based TCP implementations under many conditions, and that it provides valuable insights into the behavior of the Internet under high load from diverse user communities.