Re: Minstrel rate-setting question

[Felix Fietkau <nbd@openwrt.org>]

Larry Finger wrote:
> In the patch I just submitted, I added TX retry info to the driver for RTL8187L devices.
> One of the testers of the patch reports that the rate is being adjusted, but that the rate
> is overestimated and he has to limit it to 11M. His minstrel statistics and his comments
> are as follows:
> 
>> rate     throughput  ewma prob   this prob  this succ/attempt   success
>> attempts
>>      1         0.2       29.8       12.5          1(  8)         51
>> 86
>>      2         0.4       22.6       14.2          0(  0)          6
>> 25
>>   P  5.5       2.7       54.2      100.0          0(  0)         41
>> 109
>>     11         1.9       20.5       12.5          0(  0)          6
>> 30
>>      6         2.1       38.4       20.0          0(  0)          7
>> 24
>>      9         2.0       24.8        8.3          0(  0)          6
>> 25
>>     12         2.4       21.9       12.5          0(  0)          6
>> 52
>>     18         3.2       19.9       12.5          0(  0)          6
>> 44
>>     24         4.7       22.2       12.5          0(  0)         15
>> 109
>>     36         4.4       14.5        9.0          0(  0)        495
>> 2036
>>  t  48         5.2       13.4       12.5          0(  0)        209
>> 1220
>> T   54         8.6       19.8       14.2          1(  7)        556
>> 4514
>> 
>> But It still is a little to optimistic. But I supose is the algorithm
>> fault. It gets a little laggy and I have to limit the rate to 11Mbps.
>> 
>> If I do that the conection works perfect.
>> 
>> But the algorithm is working, maybe its just how it should work.
>> What do you think?
> 
> The losses for all rates seem to be a little high, and I plan to see if he can change
> channels, but I would like an expert's analysis of these data.
I believe that it's not minstrel's fault. According to the information that
I have about RTL8187L, the chip implements rate fallback. If the chip changes
rates during retransmissions, the status information the way you gather it
will not be accurate and will trick minstrel into believing that higher rates
actually work (which in fact they don't). 
Please do some transmissions at 54M and run a monitor mode capture on a different
card to see if it changes the rate during transmission and how frequently it
uses one rate before falling back (if indeed my theory is correct).
This information could be used to provide some more accurate feedback through
multi-rate retry status information.

- Felix