This page describes the minstrel rate control algorithm for mac80211.
minstrel is a mac80211 rate control algorithm ported over from MadWifi which supports multiple rate retries and claimed to be one of the best, if not the best, rate control algorithm.
Why did we write it ?
I had two identical nodes in my office, both running some rate algorithm. The rate algorithm was picking 6 MBits/sec as the optimal rate. Yet, if I took those nodes and used a fixed rate setting, adjusted the fixed rate until the ideal combination was found, a much better throughput (12 MBits/sec) could be achieved. Clearly, there was room for improvement.
A second test was to take those two nodes, with some rate algorithm, and look at the throughput. One node was moved behind a pillar. The measured throughput dropped. Then the nodes were moved so that the beam path was not blocked. But, the measured throughput did not increase. This test said that the rate algorithm is not responding to changes in the environment. Surely, something better can be done??
Inspection of the code in different rate algorithms left us bewildered. Why did all the code bases we looked at contain the assumption that packets sent at slow data rates are more likely to succeed than packets sent at higher datarates? The physics behind this assumption baffled us. A slow data rate packet has the highest possibility of being "shot down" by some other node sending a packet.
Thus, the decision was made to write an automatic rate algorithm that
- relied solely on measured performance with the minimum of assumptions
- responded to changes in the environment - so it must auto update correctly
- compare the performance of our algorithm at all times with that obtained by adjusting the fixed rate setting
History of minstrel
The minstrel rate control algorithm present on mac80211 was ported from MadWifi by Felix Fietkau. MadWifi's minstrel implementation was released on January 2005, originally designed and implemented by Derek Smithies Ph.D. Read minstrel's madwifi documentation.
Theory of operation
We defined the measure of successfulness (of packet transmission) as
Mega bits transmitted Prob_success_transmission * ----------------------- elapsed time
This measure of successfulness will therefore adjust the transmission speed to get the maximum number of data bits through the radio interface. Further, it means that the 1 Mbps rate (which has a very high probability of successful transmission) will not be used in preference to the 11 Mbps rate.
We decided that the module should record the successfulness of all packets that are transmitted. From this data, the module has sufficient information to decide which packets are more successful than others. However, a variability element was required. We had to force the module to examine bit rates other than optimal. Consequently, some percent of the packets have to be sent at rates regarded as non optimal.
10 times a second (this frequency is alterable by changing the driver code) a timer fires, which evaluates the statistics table. EWMA calculations (described below) are used to process the success history of each rate. On completion of the calculation, a decision is made as to the rate which has the best throughput, second best throughput, and highest probability of success. This data is used for populating the retry chain during the next 100 ms.
As stated above, the minstrel algorithm collects statistics from all packet attempts. Minstrel spends a particular percentage of frames, doing "look around" i.e. randomly trying other rates, to gather statistics. The percentage of "look around" frames defaults to 10%. The distribution of lookaround frames is also randomized somewhat to avoid any potential "strobing" of lookaround between similar nodes.
One of our test networks consisted of three nodes, A, B, C. Node C snarfed all the packets on the air for analysis in wireshark. With this network, it was possible to verify that A and B were correctly choosing the optimal rate (by manually looking at the packet dump on C). Interestingly, we saw that sometimes there was a 100ms delay between the receipt of an 802.11 ack packet and the transmission of the next data packet. Our view is that the TCP congestion control mechanism was the cause. Since the packet took so long to transmit, the TCP control code in the kernel was delaying the next data packet. Adding code so that the retry chain never exceeded 26 ms in length gave a significant increase in measured throughput.
Several devices have built in multirate retry chains. For example Atheros 802.11abg chipsets have four segments. Each segment is an advisement to the hardware to try to send the current packet at some rate, with a fixed number of retry attempts. Once the packet is successfully transmitted, the remainder of the retry chain is ignored. Selection of the number of retry attempts was based on the desire to get the packet out in under 26 ms, or fail.
There is some room for movement here - if the traffic is UDP then the limit of 26 ms for the retry chain length is "meaningless". However, one may argue that if the packet was not transmitted after some time period, it should fail. Further, one does expect UDP packets to fail in transmission. We leave it as an area for future improvement.
The (re)try segment chain is calculated in two possible manners. If this packet is a normal tranmission packet (90% of packets are this) then the retry count is best throughput, next best throughput, best probability, lowest baserate. If it is a sample packet (10% of packets are this), then the retry chain is random lookaround, best throughput, best probability, lowest base rate. In tabular format:
Try | Lookaround rate | Normal rate ------------------------------------------------ 1 | Random lookaround | Best throughput 2 | Best throughput | Next best throughput 3 | Best probability | Best probability 4 | Lowest Baserate | Lowest Baserate
The retry count is adjusted so that the transmission time for that section of the retry chain is less than 26 ms.
We have adjusted the code so that the lowest rate is never used for the lookaround packet. Our view is that since this rate is used for management packets, this rate must be working. Alternatively, the link is set up with management packets, data packets are acknowledged with management packets. Should the lowest rate stop working, the link is going to die reasonably soon.
Analysis of information showed that the system was sampling too hard at some rates. For those rates that never work (54mb, 500m range) there is no point in sending 10 sample packets (< 6 ms time). Consequently, for the very very low probability rates, we sample at most twice.
The retry chain above does "work", but performance is suboptimal. The key problem being that when the link is good, too much time is spent sampling the slower rates. Thus, for two nodes adjacent to each other, the throughput between them was several Mbps below using a fixed rate. The view was that minstrel should not sample at the slower rates if the link is doing well. However, if the link deteriorates, minstrel should immediately sample at the lower rates.
Some time later, we realized that the only way to code this reliably was to use the retry chain as the method of determining if the slower rates are sampled. The retry chain was modified as:
Try | Lookaround rate | Normal rate | random < best | random > best | -------------------------------------------------------------- 1 | Best throughput | Random rate | Best throughput 2 | Random rate | Best throughput | Next best throughput 3 | Best probability | Best probability | Best probability 4 | Lowest Baserate | Lowest baserate | Lowest baserate
With this retry chain, if the randomly selected rate is slower than the current best throughput, the randomly selected rate is placed second in the chain. If the link is not good, then there will be data collected at the randomly selected rate. Thus, if the best throughput rate is currently 54 Mbps, the only time slower rates are sampled is when a packet fails in transmission. Consequently, if the link is ideal, all packets will be sent at the full rate of 54 Mbps. Which is good.
The core of the Minstrel rate algorithm is the EWMA, or Exponential Weighted Moving Average. The EWMA is defined on wikipedia here. Using an EWMA allows us to put more importance on recent results, than older results. Consequently, we can cope with environmental changes, as old results (from a potentially different environment) are ignored.
At the beginning of this document, we described a test with two nodes, and one node was moved behind a pillar (blocking the beam) and then moving the node out from behind the pillar. For an automatic rate control algorithm, some method is required to assign more importance to recent results than old results. By using EWMA, we can achieve this. Old results have minimal impact on the choice of ideal rate.
The EWMA calculation is carried out 10 times a second, and is run for each rate. This calculation has a smoothing effect, so that new results have a reasonable (but not large) influence on the selected rate. However, with time, a series of new results in some particular direction will predominate. Given this smoothing, we can use words like inertia to describe the EWMA.
By "new results", we mean the results collected in the just completed 100 ms interval. Old results are the EWMA scaling values from before the just completed 100 ms interval.
If no packets have been sent for a particular rate in a time interval, no calculation is carried out.
The appropriate update interval was selected on the basis of choosing a compromise between
- collecting enough success/failure information to be meaningful
- minimizing the amount of cpu time spent do the updates
- providing a means to recover quickly enough from a bad rate selection.
The first two points are self explanatory. When there is a sudden change in the radio environment, an update interval of 100 ms will mean that the rates marked as optimal are very quickly marked as poor. Consequently, the sudden change in radio environment will mean that minstrel will very quickly switch to a better rate.
A sudden change in the transmission probabilities will happen when the node has not transmitted any data for a while, and during that time the environment has changed. On starting to transmit, the probability of success at each rate will be quite different. The driver must adapt as quickly as possible, so as to not upset the higher TCP network layers.