Time-based Fairness Improves Performance in Multi-rate WLANs

This paper asks a practical but fundamental question about fairness in multi-rate IEEE 802.11 WLANs: when stations use different PHY data rates (e.g., 1, 2, 5.5, 11 Mbps in 802.11b) due to heterogeneous indoor channel conditions, does the “fairness” implemented by the standard MAC (DCF) and typical AP queuing lead to efficient overall performance? The authors argue that the prevailing notion of fairness in 802.11—implicitly induced by DCF’s equal long-term transmission opportunities and by AP scheduling—tends to be throughput-based. That is, nodes with similar conditions and equal packet sizes get similar achieved throughputs, even though the faster nodes transmit frames more quickly and therefore consume less channel time per frame. In congested, rate-diverse scenarios, this throughput-based fairness can drastically reduce aggregate throughput because the channel is effectively dominated by the slower rate’s longer transmission/ACK/retransmission cycles.

The significance is twofold. First, it provides an explanation for a well-known “performance anomaly” in multi-rate WLANs: the aggregate throughput can be far below what one would expect by averaging the rates of competing clients. Second, it reframes fairness as a time-allocation problem rather than a throughput-equality problem. The authors propose time-based fairness, where each competing node receives an equal share of the wireless channel occupancy time (the fraction of time the node is able to access the channel for transmitting or receiving with the AP). This is motivated by the observation that in infrastructure WLANs the AP is a facilitator and the relevant shared resource is the channel time consumed by each client’s uplink and downlink traffic.

Methodologically, the paper combines (i) an analytic framework, (ii) validation against measurements from a deployed 802.11b setup, (iii) trace analysis to show rate diversity is common, and (iv) an implementation and experimental evaluation of a new mechanism. Analytically, the authors model a set of competing nodes with data rates and packet sizes. They define channel occupancy time for node i as a fraction of time, with the constraint $$\sum _{i\in I}T(i)=1.$$ They express each node’s achieved throughput as $$R(i)=T(i)\times \gamma (d_i,s_i,I),$$ where $\gamma$ is a “baseline throughput” that depends on the node’s rate and packet size (and implicitly on loss characteristics), representing the maximum total throughput when all nodes use the same $d$ and $s$. Under DCF, they derive a specific form for $T(i)$ when packet sizes are equal, showing that throughput-based fairness emerges: slower nodes occupy the channel longer, but the resulting throughput per node becomes equal.

The paper then contrasts this with time-based fairness. Under time-based fairness, each node gets equal channel occupancy time, $$T^{\prime }(i)=\frac{1}{n},$$ which yields $$R^{\prime }(i)=\frac{{\gamma }_i}{n}.$$ A key property proved/argued is the “baseline property”: a node’s throughput under time-based fairness equals the throughput it would achieve in a single-rate WLAN where all nodes transmit at its own rate. This property is central to the paper’s claim that time-based fairness improves aggregate performance without making any node worse than in the corresponding single-rate baseline.

For empirical evaluation, the authors implement their scheme, TBR (Time-based Regulator), in an AP driver (HostAP on Linux) and run experiments with an AP (D-Link DWL-650 with HostAP) and iPAQs equipped with Cisco-350 cards. Each experiment sends about 2000 packets of size 1500 bytes per node, and each data point is averaged over 5–10 runs. They compare a baseline configuration (Exp-Normal, no TBR) against Exp-TBR (TBR enabled). They evaluate both correctness (matching the analytic equations) and performance (throughput gains) in both downlink and uplink.

Key findings are presented in multiple forms. First, the authors show analytically and experimentally that DCF’s throughput-based fairness can reduce aggregate throughput when nodes use different rates. In their motivating example (two TCP uploaders), the faster node’s throughput can drop by about 3.75× when competing against a 1 Mbps node rather than another 11 Mbps node, and the aggregate throughput is far below naive expectations (they state 1.34 Mbps vs an expected 2.93 Mbps in the 11 Mbps vs 1 Mbps case). Second, they provide a concrete comparison table for four nodes transmitting at 1, 2, 11, and 11 Mbps. Under throughput-based fairness (RF), each node achieves 0.436 Mbps and the total is 1.742 Mbps; under time-based fairness (TF), the throughputs are 0.202, 0.373, 1.30, and 1.30 Mbps with a total of 3.175 Mbps. This corresponds to a large aggregate improvement (about 82% relative to RF in that scenario).

Third, they validate their analytic model against experiments. In downlink TCP experiments with two nodes using different rates, they report that when TBR is used, total achieved throughput increases by about 6% in the 5.5 vs 11 Mbps case, 35% in the 2 vs 11 Mbps case, and 103% in the 1 vs 11 Mbps case. They also show that Exp-TBR results closely match the predictions from the time-based fairness equation (Eq. 12), with small discrepancies attributed to TBR’s estimation of channel occupancy time without per-packet retransmission information.

Fourth, they show that TBR works in uplink as well, reporting similar improvements in the uplink direction (Figure 9(b) mirrors the downlink pattern). They also test a scenario with mixed bottleneck capacities: two nodes both transmit at 11 Mbps, but one node’s bottleneck is limited to 2.1 Mbps. They report no significant difference between Exp-Normal and Exp-TBR in this case (Table 4: n1 is 2.9434 Mbps vs 2.9542 Mbps; n2 is 2.1276 Mbps vs 2.1193 Mbps; total 5.071 vs 5.061 Mbps), supporting the claim that TBR’s rate adjustment preserves appropriate fairness when demand differs.

The paper also includes evidence that rate diversity is common. Using packet traces from a 90-minute MIT student workshop, they find that during WS-2 more than 30% of bytes were transferred at rates lower than 11 Mbps. In an office experiment (EXP-1), more than 50% of bytes were transferred at the lowest data rate when receivers were placed at varying distances and through walls. They further analyze a Dartmouth residential facility trace (Kotz et al.) to argue that during busy intervals, the heaviest user rarely monopolizes the channel; other users contribute significant throughput during congested 1-second intervals.

Limitations are acknowledged both explicitly and implicitly. Analytically, the model assumes a fluid traffic setting for aggregate throughput and abstracts away varying loss characteristics by focusing on how rate diversity affects baseline throughput $\gamma$. In the implementation, TBR’s computation of channel occupancy time in uplink requires retransmission information; when such information is not available, the AP estimates it, which the authors note can bias channel time accounting (slightly decreasing total throughput relative to the ideal equation). They also acknowledge that TBR’s responsiveness depends on token-rate adjustment frequency and the burst allowance parameter $b$ (${\text{bucket}}_i$), and that short-lived flows may not receive fair shares quickly enough. Finally, they note that their current TBR implementation does not include client-side cooperation for all cases; while experiments show that long-term guarantees hold for TCP flows even without client changes, correctness for significant uplink UDP fractions may require additional client logic.

Practically, the results matter for anyone operating or designing AP-based WLANs, especially in environments where 802.11g and 802.11b clients coexist or where indoor channel heterogeneity is high. Network operators and WLAN vendors should care because time-based fairness can substantially improve aggregate throughput during congestion while preserving a per-node baseline guarantee tied to single-rate performance. The proposed TBR mechanism is designed to be backward compatible at the MAC level (implemented as an AP driver change) and can be deployed incrementally. Researchers should care because the paper provides a clear fairness redefinition and a quantitative framework linking fairness notions to throughput and task-completion metrics.

Overall, the paper’s core contribution is the argument and demonstration that in multi-rate WLANs, fairness should be defined in terms of channel occupancy time rather than throughput. By implementing time-based fairness via TBR, the authors show large aggregate throughput gains (up to 103% in their 1 vs 11 Mbps downlink experiments) while maintaining a baseline property that prevents slower nodes from being arbitrarily harmed relative to their single-rate counterparts.