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
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Link: https://lore.kernel.org/r/20191008171139.96476-2-nbd@nbd.name
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
On some devices that only support static rate fallback tables sending rate
control probing packets can be really expensive.
Probing lower rates can already hurt throughput quite a bit. What hurts even
more is the fact that on mt76x0/mt76x2, single probing packets can only be
forced by directing packets at a different internal hardware queue, which
causes some heavy reordering and extra latency.
The reordering issue is mainly problematic while pushing lots of packets to
a particular station. If there is little activity, the overhead of probing is
neglegible.
The static fallback behavior is designed to pretty much only handle rate
control algorithms that use only a very limited set of rates on which the
algorithm switches up/down based on packet error rate.
In order to better support that kind of hardware, this patch implements a
different approach to rate probing where it switches to a slightly higher rate,
waits for tx status feedback, then updates the stats and switches back to
the new max throughput rate. This only triggers above a packet rate of 100
per stats interval (~50ms).
For that kind of probing, the code has to reduce the set of probing rates
a lot more compared to single packet probing, so it uses only one packet
per MCS group which is either slightly faster, or as close as possible to
the max throughput rate.
This allows switching between similar rates with different numbers of
streams. The algorithm assumes that the hardware will work its way lower
within an MCS group in case of retransmissions, so that lower rates don't
have to be probed by the high packets per second rate probing code.
To further reduce the search space, it also does not probe rates with lower
channel bandwidth than the max throughput rate.
At the moment, these changes will only affect mt76x0/mt76x2.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Link: https://lore.kernel.org/r/20190820095449.45255-4-nbd@nbd.name
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Based on 2 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license version 2 as
published by the free software foundation
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license version 2 as
published by the free software foundation #
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-only
has been chosen to replace the boilerplate/reference in 4122 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Enrico Weigelt <info@metux.net>
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190604081206.933168790@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
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@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Legacy-only devices are not very common and the overhead of the extra
code for HT and VHT rates is not big enough to justify all those extra
lines of code to make it optional.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
debugfs entries are cleaned up by debugfs_remove_recursive already.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This avoids the costly int_sqrt calls in the statistics update and moves
it to the debugfs code instead.
This also fixes an overflow in the previous standard deviation
calculation.
Signed-off-by: Thomas Huehn <thomas.huehn@evernet-eg.de>
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
The loss of a bit of extra precision does not hurt the calculation, 12
bits is still enough to calculate probabilities well. Reducing the scale
makes it easier to avoid overflows
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This field is redundant, because it is simply last success divided by
last attempt count. Removing it from the rate stats struct saves about
1.2 KiB per HT station.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
They are only used for debugging purposes and take a very long time to
overflow. Visibly reduces the size of the per-sta rate control data.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This patch adds the statistical descriptor "standard deviation"
to better describe the current properties of Minstrel and
Minstrel-HTs success probability distribution. The standard
deviation (SD) is calculated as exponential weighted moving
standard deviation (EWMSD) and its current value is added as
new column in all rc_stats (in debugfs).
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This patch reduces the calculation costs of the EWMA macro from
"2x multiplication and 1 addition" down to "1x multiplication and
2x additions". This slightly improves performance depending on the
CPU architecture.
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This patch adds the new statistic "maximum possible lossless
throughput" to Minstrels and Minstrel-HTs rc_stats (in debugfs). This
enables comprehensive comparison between current per-rate throughput
and max. achievable per-rate throughput.
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This patch moves Minstrels and Minstrel-HTs per-rate throughput
calculation (EWMA(thr)) into a dedicated function to be called.
Therefore the variable "unsigned int cur_tp" within struct
"minstrel_rate_stats" becomes obsolete. and is removed to free
up its space.
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This patch ensures a consistent usage of variable names for type
"minstrel_rate_stats" to be used as "mrs" and from type minstrel_rate
as "mr" across both Minstrel & Minstrel-HT. In addition some
variable and function names got changed to more meaningful ones.
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This patch unifies the calculation of Minstrels and Minstrel-HTs
per-rate statistic. The new common function minstrel_calc_rate_stats()
is called when a statistic update is performed.
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This patch adds a new debugfs file "rc_stats_csv" to output Minstrels
statistics in a common csv format that is easy to parse.
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Signed-off-by: Stefan Venz <ikstream86@gmail.com>
Acked-by: Felix Fietkau <nbd@openwrt.org>
[remove printing current time of day]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
On minstrel_ht, the size of the per-sta struct is almost 18k, making it
an order-3 allocation.
A few fields inside the per-rate statistics are bigger than they need to
be. This patch reduces the size enough to cut down the per-sta struct to
about 13k (order-2 allocation).
Signed-off-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Minstrel and Mintrel_HT used there own structs to keep track of rate
statistics. Unify those variables in struct minstrel_rate_states and
move it to rc80211_minstrel.h for common usage. This is a clean-up
patch to prepare Minstrel and Minstrel_HT codebase for upcoming TPC.
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Change the code to allow making all the rate control ops
const, nothing ever needs to change them. Also change all
drivers to make use of this and mark the ops const.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Pass the rate selection table to mac80211 from minstrel_update_stats.
Only rates for sample attempts are set in info->control.rates, with deferred
sampling, only the second slot gets changed.
Signed-off-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Use powers of two in ewma of minstrel.
This changes :
- EWMA_DIV from 100 to 2^7
- EWMA_LEVEL from 75 (/EWMA_DIV=100) to 2^6 + 2^5 (/EWMA_DIV=128)
Note that this changes EWMA_DIV - EWMA_LEVEL from 25 to 2^5 and keeps
EWMA_LEVEL / EWMA_DIV == 0.75.
Signed-off-by: Karl Beldan <karl.beldan@rivierawaves.com>
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This patch improves the way minstrel sorts rates according to throughput
and success probability. 3 FOR-loops across the entire rate set in function
minstrel_update_stats() which where used to determine the fastest, second
fastest and most robust rate are reduced to 1 FOR-loop.
The sorted list of rates according throughput is extended to the best four
rates as we need them in upcoming joint rate and power control. The sorting
is done via the new function minstrel_sort_best_tp_rates().
The most robust rate selection is aligned with minstrel_ht's approach.
Once any success probability is above 95% the one with the highest
throughput is chosen as most robust rate. If success probabilities of all
rates are below 95%, the rate with the highest succ. prob. is elected as
most robust one
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
While minstrel bootstraps and fills the success probabilities of each
rate the lowest rate has typically a very high success probability
(often 100% in our tests).
Its statistics are never updated but considered to setup the mrr chain.
In our tests we see that especially the 3rd mrr stage (which is that
rate providing highest success probability) is filled with the lowest rate
because its initial high sucess probability is never updated. By design
the 4th mrr stage is filled with the lowest rate so often 3rd and 4th
mrr stage are equal.
This patch follows minstrels general approach of assuming as little
as possible about rate dependencies. Consequently we include the
lowest rate into the random sampling table to get balanced up-to-date
statistics of all rates and therefore balanced decisions.
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Minstrel's decision which rate should be directly sampled within the
1st mrr stage is limited to such rates faster than the current max
throughput rate. All rates below the current max. throughput rate
are indirectly sampled via the 2nd mrr stage.
This approach leads to deprecated per rate statistics and therfore
a deprecated mrr chain setup.
This patch uses the sampling approach from minstrel_ht. A counter is
added to sum all indirect sample attempts per rate. After 20 indirect
sampling attempts the rate is directly sampled within the 1st mrr stage.
Therefore more up-to-date statistics for all rates are maintained and
used to setup the mrr chain.
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Add documentation and more verbose variable names to minstrel's
multi-rate-retry setup within function minstrel_get_rate() to
increase the readability of the algorithm.
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Both minstrel versions use individual ways to scale up integer values
to perform calculations. Merge minstrel_ht's scaling macros into
minstrels header file and use them in both minstrel versions.
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Both rate control algorithms (minstrel and minstrel_ht) calculate
averages based on EWMA. Shift function minstrel_ewma() into
rc80211_minstrel.h and make use of it in both minstrel version.
Also shift the default EWMA level (75%) definition to the header file
and clean up variable usage.
Acked-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Thomas Huehn <thomas@net.t-labs.tu-berlin.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
When MCS rates start to get bad in 2.4 GHz because of long range or
strong interference, CCK rates can be a lot more robust.
This patch adds a pseudo MCS group containing CCK rates (long preamble
in the lower 4 slots, short preamble in the upper slots).
Signed-off-by: Felix Fietkau <nbd@openwrt.org>
[make minstrel_ht_get_stats static]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Found several threads about fixed rate mode in minstrel_ht for test
environments, but no patches for it.
This patch provides such a mode through debugfs.
Signed-off-by: John W. Linville <linville@tuxdriver.com>
This patch makes it possible to reuse the minstrel rate control ops
from another rate control module. This is useful in preparing for the
new 802.11n implementation of minstrel, which will reuse the old code
for legacy stations.
Signed-off-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
This patch cleans up the debugfs read function for the statistics by
using simple_read_from_buffer instead of its own semi-broken hack.
Also removes a useless member of the minstrel debugfs info struct.
Signed-off-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
I noticed that the throughput field of the minstrel_rate struct is never used,
so remove it.
Signed-off-by: Arnd Hannemann <hannemann@nets.rwth-aachen.de>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
This patch enhances minstrel's performance for non-MRR setups,
by preventing it from sampling slower rates with >95% success
probability and by putting at least 1 non-sample frame between
several sample frames.
Signed-off-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
