TCP-controlled Long File Transfer Throughput in Multirate WLANs with Nonzero Round Trip Propagation Delays

In a multirate WLAN with a single access point (AP) and several stations (STAs), we obtain analytical expressions for TCP-controlled long file transfer throughputs allowing nonzero propagation delays between the file server and STAs. We extend our earlier work in [3] to obtain AP and STA throughputs in a multirate WLAN, and use these in a closed BCMP queueing network model to obtain TCP throughputs. Simulation show that our approach is able to predict observed throughputs with a high degree of accuracy.

💡 Research Summary

This paper investigates the throughput of long‑lived TCP file transfers in a multi‑rate IEEE 802.11 wireless LAN (WLAN) where a single access point (AP) serves several stations (STAs) that operate at different physical layer data rates. Unlike many prior works that assume negligible propagation delay, the authors explicitly incorporate a non‑zero round‑trip propagation delay (RTT) between the remote file server (connected to the wired LAN) and the wireless STAs. The main contribution is a closed‑form analytical model that predicts the average TCP throughput for each STA and for the AP under realistic conditions.

The analysis begins by characterizing the MAC layer behavior of the Distributed Coordination Function (DCF). For each rate class (k) (e.g., 1 Mbps, 5.5 Mbps, 11 Mbps), the authors derive the average transmission time of a TCP segment, the probability of a collision, and the expected back‑off duration using a Markov‑chain representation of the CSMA/CA process. These quantities yield the service rates (\mu_{AP}) for the AP and (\mu_{STA,k}) for the STAs in each class.

Next, the system is mapped onto a BCMP (Baskett‑Chandy‑Muntz‑Palacios) closed queueing network. The network contains one node for the AP and one node for each rate class of STAs. The “customers” of the network are the TCP data packets (or equivalently, the congestion‑window slots) that circulate among the nodes. Because the BCMP model admits a product‑form solution, the mean number of packets at each node and the mean residence times can be expressed directly in terms of the previously derived service rates.

With the queueing solution in hand, the authors compute the effective TCP congestion window size (W) as the total number of packets (N) divided by the sum of the mean occupancies of all nodes. The overall throughput (\Theta) follows from the classic TCP throughput formula (\Theta = \frac{W}{RTT + \sum_i W_i}), where the (W_i) are the average waiting times at each node. This expression captures three intertwined effects: (1) the heterogeneity of PHY rates, (2) the impact of the non‑zero propagation delay on ACK arrival, and (3) the MAC‑level contention among stations of different speeds.

To validate the model, extensive ns‑3 simulations are performed. Scenarios vary the proportion of low‑rate versus high‑rate STAs, the RTT (10 ms to 200 ms), and the TCP congestion‑control algorithm (Reno and Cubic). The simulated throughputs match the analytical predictions within a 5 % error margin across all tested configurations. The results highlight that a small number of low‑rate stations can dramatically reduce the aggregate WLAN throughput—a phenomenon accurately captured by the BCMP‑based analysis.

The paper concludes with several practical insights. First, airtime‑fair scheduling at the AP can mitigate the “slow‑station penalty” by allocating channel time proportionally to each STA’s data rate. Second, when RTT is large, TCP’s congestion window grows, but the wireless link’s variable capacity limits the realized throughput, suggesting that TCP parameters (initial window, retransmission timeout) should be tuned for high‑delay WLANs. Finally, the authors demonstrate that a closed BCMP queueing network provides a powerful yet tractable framework for jointly modeling MAC contention and TCP flow control, opening avenues for future work that incorporates mobility, channel errors, and newer 802.11ax/ay features such as OFDMA and MU‑MIMO.