Loss Probability Estimation

So far, the potential bandwidth estimation methodology assumed no packet losses. Losses occur due to collisions when multiple wireless stations transmit simultaneously and also due to environmental effects such as multipath, fading etc. Packet losses reduce the bandwidth between communicating stations, since they cause nodes to double their contention window and thereby, backoff for longer durations before retransmitting their data.

Thus, in order to estimate the potential downstream bandwidth from a given AP, an end-host needs to estimate the loss rate on the wireless link from the AP to itself. We propose that nodes infer frame losses, by exploiting the 12-bit sequence number field present in the 802.11 data and management frames. An end-host passively monitors all frames transmitted by the AP for a certain duration. The end-host can then infer data frame losses based on gaps in sequence numbers during the monitoring period. It is possible that the monitoring node may hear a data frame from an AP that is a retransmission of an earlier frame, which it did not hear. In this case, the monitoring node can detect retransmissions by looking at the Retry bit in the Frame Control field of the received frame. If this bit is set, it indicates that the frame is a retransmission of an earlier frame. Since the Retry bit does not indicate the number of retransmissions of a frame, we make a simplifying assumption that the probability of more than two successive retransmissions of a frame between an AP and a host affiliated to that AP is negligible.

The above described method of inferring loss rate, is useful both in the presence of $RTS/CTS$ and in its absence. In the presence of $RTS/CTS$, the probability of an $RTS$ frame loss differs from the probability of a data frame loss, since an $RTS$ frame is transmitted at the base rate. An $RTS$ frame loss can be inferred by a monitoring end-host, if the monitoring host overhears a data frame transmission from an AP to an end-host, but does not hear the $RTS$ frame transmission from the AP to the end-host preceding the data transmission. Just as in the case of a data frame, an $RTS$ frame retransmission can be detected from the Retry bit in the Frame Control field of the frame. Data frame losses can be detected from the missing sequence numbers over the monitoring period.

The estimated loss probability can be used to calculate the expected delays incurred by the $RTS$ frames and data frames transmitted by an AP. For simplicity, we assume that $CTS$ and $ACK$ frames from the end-host to the AP are transmitted loss-free. This may be a reasonable assumption since $CTS$ and $ACK$ frames are very short. Furthermore, $CTS$ frames are transmitted at the base rate and the $ACK$ frames are transmitted collision-free. This assumption means that $CTS$ and $ACK$ frames always incur fixed delays. Losses then only impact the $RTS$ and data frames in our model. The estimated loss probability can easily be incorporated to obtain the expected back-off delay and the corresponding frame delay, using the analysis shown in [6].

When there are no affiliated hosts, a monitoring node does not overhear any transmissions except the beacon frames transmitted by an AP. Absence of a beacon frame in a beacon interval indicates that the beacon frame was lost. A monitoring host can estimate the loss probability of data frames to be the loss probability of the beacon frames. The $RTS$ frames are transmitted at the base rate and can be assumed to be transmitted loss-free, especially given that there is no contention for the medium and that the probability of a collision is zero.