Hi Johannes, On Sat, Jun 9, 2012 at 1:55 AM, Johannes Berg <johannes@xxxxxxxxxxxxxxxx> wrote: > On Fri, 2012-06-08 at 13:31 -0700, Jason Abele wrote: > >> + if (sta->tx_packets > 100) >> + rate = max(ewma_read(&sta->avg_rate), 10UL); >> + else >> + /* Not enough traffic sent, use rx signal strengh as proxy */ >> + rate = rssi_get_rate(sta); > > This just seems completely wrong to me. There's very little correlation > between bitrates and RSSI, in fact, we once had 0 dBm power by accident > and got better performance than with 15 dBm power (due to another bug, > but still)... I would say this is an improvement over the existing approach, which assigns the lowest rate (and therefore a poor metric) to all mesh paths over new peers. This means that a possibly good path through a new peer is not used until existing mesh paths degrade below this initial poor metric. Alternatively we had considered to initialize the metric with the best (lowest) metric but that just reverses the problem, as the new peers would attract traffic away from good established paths. By using RSSI we can at least choose a sensible initial value for the airtime link metric. I agree with you in that a high RSSI does not necessarily mean higher rates, but I would not say that there is "very little correlation" between the two. SNR is directly correlated with datarates¹. Unfortunately SNR measurements are not available at the mac80211 layer. But at equal noise levels, which is not unlikely between two direct peer neighbors, RSSI will correlate with SNR measurements. Keep in mind that the RSSI is only used until traffic starts flowing through that peer. Once valid rate statistics are available, the RSSI is no longer used to calculate the metric. > Looking at the code that uses this, I also note: > >> /* This should be adjusted for each device */ >> int device_constant = 1 << ARITH_SHIFT; > > which seems *totally* wrong. Maybe it's time to rethink this code? Indeed, but the standard defines this constant that "Varies depending on PHY" (see sect. 13.9). We would have to get each driver to report the right value. Cheers, Javier [1] For instance, see http://www.inf.ed.ac.uk/teaching/courses/cn/papers/ygz-infocom08.pdf > johannes > > _______________________________________________ > Devel mailing list > Devel@xxxxxxxxxxxxxxxxxxxx > http://lists.open80211s.org/cgi-bin/mailman/listinfo/devel -- Javier Cardona cozybit Inc. http://www.cozybit.com -- To unsubscribe from this list: send the line "unsubscribe linux-wireless" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
