On 2018-10-06 19:59, Dave Taht wrote:
> On Sat, Oct 6, 2018 at 10:37 AM Felix Fietkau <nbd@xxxxxxxx> wrote:
>>
>> When there are few packets (e.g. for sampling attempts), the exponentially
>> weighted variance is usually vastly overestimated, making the resulting data
>> essentially useless. As far as I know, there has not been any practical use
>> for this, so let's not waste any cycles on it.
>>
>> Signed-off-by: Felix Fietkau <nbd@xxxxxxxx>
>> ---
>> net/mac80211/rc80211_minstrel.c | 6 -----
>> net/mac80211/rc80211_minstrel.h | 26 +---------------------
>> net/mac80211/rc80211_minstrel_debugfs.c | 14 ++++--------
>> net/mac80211/rc80211_minstrel_ht_debugfs.c | 14 ++++--------
>> 4 files changed, 9 insertions(+), 51 deletions(-)
>>
>> diff --git a/net/mac80211/rc80211_minstrel.c b/net/mac80211/rc80211_minstrel.c
>> index dead57ba9eac..a34e9c2ca626 100644
>> --- a/net/mac80211/rc80211_minstrel.c
>> +++ b/net/mac80211/rc80211_minstrel.c
>> @@ -167,12 +167,6 @@ minstrel_calc_rate_stats(struct minstrel_rate_stats *mrs)
>> if (unlikely(!mrs->att_hist)) {
>> mrs->prob_ewma = cur_prob;
>> } else {
>> - /* update exponential weighted moving variance */
>> - mrs->prob_ewmv = minstrel_ewmv(mrs->prob_ewmv,
>> - cur_prob,
>> - mrs->prob_ewma,
>> - EWMA_LEVEL);
>> -
>> /*update exponential weighted moving avarage */
>> mrs->prob_ewma = minstrel_ewma(mrs->prob_ewma,
>> cur_prob,
>> diff --git a/net/mac80211/rc80211_minstrel.h b/net/mac80211/rc80211_minstrel.h
>> index 54b2b2c3e10a..23ec953e3a24 100644
>> --- a/net/mac80211/rc80211_minstrel.h
>> +++ b/net/mac80211/rc80211_minstrel.h
>> @@ -35,19 +35,6 @@ minstrel_ewma(int old, int new, int weight)
>> return old + incr;
>> }
>>
>> -/*
>> - * Perform EWMV (Exponentially Weighted Moving Variance) calculation
>> - */
>
> I worry about this one. where are you getting your proof from?
I've done quite a few measurements myself to see if this can be usable
for further rate control improvements or for the upcoming TPC work.
The data this generates simply fluctuates wildly and incoherently based
on the sampling behavior, making it completely useless.
Together with Thomas (who introduced this code), I tried a few times to
fix this, but couldn't find any way to make it coherent and usable.
Thomas and I both agreed that it's better to just remove it until
somebody has a better idea what to do.
Also, this was only used for debugfs statistics, not for any actual rate
control behavior.
- Felix
