Minstrel Introduction ================================================================== This code is called minstrel, because we have taken a wandering minstrel approach. Wander around the different rates, singing wherever you can. And then, look at the performance, and make a choice. Note that the wandering minstrel will always wander in directions where he/she feels he/she will get paid the best for his/her work. The Minstrel autorate selection algorithm is an EWMA based algorithm and is derived from 1)An initial rate module we released in 2005, http://sourceforge.net/mailarchive/forum.php?forum_id=33966&max_rows=25&style=flat&viewmonth=200501&viewday=5 2)the "sample" module in the madwifi-ng source tree. The code released in 2005 had some algorithmic and implementation flaws (one of which was that it was based on the old madwifi codebase) and was shown to be unstable. Performance of the sample module is poor (http://www.madwifi.org/ticket/989), and we have observed similar issues. We noted: 1)The rate chosen by sample did not alter to match changes in the radio environment. 2)Higher throughput (between two nodes) could often be achieved by fixing the bitrate of both nodes to some value. 3)After a long period of operation, "sample" appeared to be stuck in a low data rate, and would not move to a higher data rate. We examined the code in sample, and decided the best approach was a rewrite based on sample and the module we released in January 2005. 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 1mb/sec rate (which has a very high probability of successful transmission) will not be used in preference to the 11mb/sec 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, is set at boot time via the module parameter "ath_lookaround_rate" and defaults to 10%. The distribution of lookaround frames is also randomised somewhat to avoid any potential "strobing" of lookaround between similar nodes. TCP theory tells us that each packet sent must be delivered in under 26ms. Any longer duration, and the TCP network layers will start to back off. A delay of 26ms implies that there is congestion in the network, and that fewer packets should be injected to the device. Our conclusion was to adjust the retry chain of each packet so the retry chain was guaranteed to be finished in under 26ms. Retry Chain ================================================================== The HAL provides a multirate retry chain - which consists of four segments. Each segment is an advisement to the HAL 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 26ms, or fail. We provided a module parameter, ath_segment_size, which has units of microseconds, and specifies the maximum duration one segment in the retry chain can last. This module parameter has a default of 6000. Our view is that a segment size of between 4000 and 6000 seems to fit most situations. There is some room for movement here - if the traffic is UDP then the limit of 26ms 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. After some discussion, 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 in the /proc/net/madwifi/athX/rate_info file 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 (<6ms 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 megabits/sec 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 realised 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 54mbs, 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 54mbs. Which is good. EWMA ================================================================== 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 100ms interval. Old results are the EWMA scaling values from before the just completed 100ms interval. EWMA scaling is set by the module parameter ath_ewma_level, and defaults to 75%. A value of 0% means use only the new results, ignore the old results. A value of 99% means use the old results, with a tiny influence from the new results. The calculation (performed for each rate, at each timer interrupt) of the probability of success is: Psucces_this_time_interval * (100 - ath_ewma_level) + (Pold * ath_ewma_level) Pnew = ------------------------------------------------------------------------------ 100 number_packets_sent_successfully_this_rate_time_interval Psuccess_this_time_interval=-------------------------------------------------------- number_packets_sent_this_rate_time_interval If no packets have been sent for a particular rate in a time interval, no calculation is carried out. The Psuccess value for this rate is not changed. If the new time interval is the first time interval (the module has just been inserted), then Pnew is calculated from above with Pold set to 0. The appropriate update interval was selected on the basis of choosing a compromise between *collecting enough success/failure information to be meaningful *minimising 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 100ms 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. Module Parameters ==================================================== The module has three parameters: name default value purpose ath_ewma_level 75% rate of response to new data. A value of 100 is VERY responsive. ath_lookaround_rate 10% percent of packets sent at non optimal speed. ath_segment_size 6000 maximum duration of a retry segment (microseconds) Test Network ==================================================== We used three computers in our test network. The first two, equipped with atheros cards running in adhoc mode. We used a program that sends a fixed number of TCP packets between computers, and reports on the data rate. The application reports on the data rate - at an application layer level, which is considerably lower than the level used in transmitting the packets. The third computer had an atheros card also, but was running network monitor mode with ethereal. The monitor computer was used to see what data rates were used on the air. This computer was a form of "logging of the connection" without introducing any additional load on the first two computers. It was from monitoring the results on the third computer that we started to get some confidence in the correctness of the code. We observed TCP backoffs (described above) on this box. There was much celebration when the throughput increased simply because the retry chain was finished in under 26 ms. Our view was that throughput between the two computers should be as close as possible (or better than) what can be achieved by setting both ends to fixed rates. Thus, if setting the both ends to fixed rates significantly increases the throughput, a reasonable conclusion is that a fault exists in the adaptive rate rate. We recorded throughputs (with minstrel) that are within 10% of what is achieved with the experimentally determined optimum fixed rate. Notes on Timing ==================================================== As noted above, Minstrel calculates the throughput for each rate. This calculation (using a packet of size 1200 bytes) determines the transmission time on the radio medium. In these calculations, we assume a contention window min and max value of 4 and 10 microseconds respectively. Further, calculation of the transmission time is required so that we can guarantee a packet is transmitted (or dropped) in a minimum time period. The transmission time is used in determining how many times a packet is transmitted in each segment of the retry chain. Indeed, the card will supply the cwmin/cwmax values directly iwpriv if_name get_cwmin <0|1|2|3> <0|1> We have not made direct calls to determine cwmin/cwmax - this is an area for future work. Indeed, the cwmin/cwmax determination code could check to see if the user has altered these values with the appropriate iwpriv. The contention window size does vary with traffic class. For example, video and voice have a contention window min of 3 and 2 microseconds respectively. Currently, minstrel does not check traffic class. Calculating the throughputs based on traffic class and bit rate and variable packet size will significantly complicate the code and require many more sample packets. More sample packets will lower the throughput achieved. Thus, our view is that for this release, we should take a simple (but reasonable) approach that works stably and gives good throughputs. Values of cwin/cwmax of 4 and 10 microseconds are from IEEE Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications, Amendment : Medium Access Control (MAC) Enhancements for Quality of Service (QoS) P802.11e/D12.0, November 2004. Internal variable reporting ==================================================== The minstrel algorithm reports to the proc file system its internal statistics, which can be viewed as text. A sample output is below: cat /proc/net/madwifi/ath0/rate_info rate data for node:: 00:02:6f:43:8c:7a rate throughput ewma prob this prob this succ/attempt success attempts 1 0.0 0.0 0.0 0( 0) 0 0 2 1.6 92.4 100.0 0( 0) 11 11 5.5 3.9 89.9 100.0 0( 0) 11 11 11 6.5 86.6 100.0 0( 0) 10 11 6 4.8 92.4 100.0 0( 0) 11 11 9 6.9 92.4 100.0 0( 0) 11 11 12 9.0 92.4 100.0 0( 0) 11 11 18 12.1 89.9 100.0 0( 0) 11 11 24 15.5 92.4 100.0 0( 0) 11 11 36 20.6 92.4 100.0 0( 0) 11 11 t 48 23.6 89.9 100.0 0( 0) 12 12 T P 54 27.1 96.2 96.2 0( 0) 996 1029 Total packet count:: ideal 2344 lookaround 261 There is a separate table for each node in the neighbor table, which will appear similar to above. The possible datarates for this node are listed in the column at the left. The calculated throughput and "ewma prob" are listed next, from which the rates used in retry chain are selected. The rates with the maximum throughput, second maximum throughput and maximum probability are indicated by the letters T, t, and P respectively. The statistics gathered in the last 100ms time period are displayed in the "this prob" and "this succ/attempt" columns. Finally, the number of packets transmitted at each rate, since module loading are listed in the last two columns. When interpreting the last four columns, note that we use the words "succ" or "success" to mean packets successfully sent from this node to the remote node. The driver determines success by analysing reports from the hal. The word "attempt" or "attempts" means the count of packets that we transmitted. Thus, the number in the success column will always be lower than the number in the attempts column. When the two nodes are brought closer together, the statistics start changing, and you see more successful attempts at the higher rates. The ewma prob at the higher rates increases and then most packets are conveyed at the higher rates. When the rate is not on auto, but fixed, this table is still available, and will report the throughput etc for the current bit rate. Changing the rate from auto to fixed to auto will completely reset this table, and the operation of the minstrel module.