On 2019-09-29 20:42, Toke Høiland-Jørgensen wrote: > Felix Fietkau <nbd@xxxxxxxx> writes: > >> Rate success probability usually fluctuates a lot under normal conditions. >> With a simple EWMA, noise and fluctuation can be reduced by increasing the >> window length, but that comes at the cost of introducing lag on sudden >> changes. >> >> This change replaces the EWMA implementation with a moving average that's >> designed to significantly reduce lag while keeping a bigger window size >> by being better at filtering out noise. >> >> It is only slightly more expensive than the simple EWMA and still avoids >> divisions in its calculation. >> >> The algorithm is adapted from an implementation intended for a completely >> different field (stock market trading), where the tradeoff of lag vs >> noise filtering is equally important. It is based on the "smoothing filter" >> from http://www.stockspotter.com/files/PredictiveIndicators.pdf. >> >> I have adapted it to fixed-point math with some constants so that it uses >> only addition, bit shifts and multiplication >> >> To better make use of the filtering and bigger window size, the update >> interval time is cut in half. >> >> For testing, the algorithm can be reverted to the older one via >> debugfs > > This looks interesting! Do you have any performance numbers from your > own testing to share? :) To show the difference, I also generated some random data, ran it through minstrel's EWMA and the new code and made a plot: http://nbd.name/ewma-filter-plot.png The real world test that I did was using mt76x2: I ran 3 iperf TCP streams from an AP to a station in a cable setup with an attenuator. I switched from 70 dB attenuation to 40 dB and measured the time it takes for TCP throughput to stabilize at a higher rate. Without my changes it takes about 5-6 seconds, with my changes it's only 2-3 seconds. - Felix
