Received-SPF: pass (google.com: domain of linux-wireless-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Toke =?utf-8?Q?H=C3=B8iland-J=C3=B8rgensen?= <toke@redhat.com>
To:     Felix Fietkau <nbd@nbd.name>, linux-wireless@vger.kernel.org
Cc:     johannes@sipsolutions.net
Subject: Re: [PATCH 4/6] mac80211: minstrel_ht: significantly redesign the
 rate probing strategy
In-Reply-To: <3ac94bcf-4c76-9c1e-c903-0810d2067b3a@nbd.name>
References: <20210124122812.49929-1-nbd@nbd.name>
 <20210124122812.49929-4-nbd@nbd.name> <87o8hdmdqs.fsf@toke.dk>
 <3ac94bcf-4c76-9c1e-c903-0810d2067b3a@nbd.name>
Date:   Mon, 25 Jan 2021 22:21:30 +0100
Message-ID: <87eei8d86d.fsf@toke.dk>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

Felix Fietkau <nbd@nbd.name> writes:

> On 2021-01-25 12:56, Toke H=C3=B8iland-J=C3=B8rgensen wrote:
>> Felix Fietkau <nbd@nbd.name> writes:
>>=20
>>> The biggest flaw in current minstrel_ht is the fact that it needs way t=
oo
>>> many probing packets to be able to quickly find the best rate.
>>> Depending on the wifi hardware and operating mode, this can significant=
ly
>>> reduce throughput when not operating at the highest available data rate.
>>>
>>> In order to be able to significantly reduce the amount of rate sampling,
>>> we need a much smarter selection of probing rates.
>>>
>>> The new approach introduced by this patch maintains a limited set of
>>> available rates to be tested during a statistics window.
>>>
>>> They are split into distinct categories:
>>> - MINSTREL_SAMPLE_TYPE_INC - incremental rate upgrade:
>>>   Pick the next rate group and find the first rate that is faster than
>>>   the current max. throughput rate
>>> - MINSTREL_SAMPLE_TYPE_JUMP - random testing of higher rates:
>>>   Pick a random rate from the next group that is faster than the current
>>>   max throughput rate. This allows faster adaptation when the link chan=
ges
>>>   significantly
>>> - MINSTREL_SAMPLE_TYPE_SLOW - test a rate between max_prob, max_tp2 and
>>>   max_tp in order to reduce the gap between them
>>>
>>> In order to prioritize sampling, every 6 attempts are split into 3x INC,
>>> 2x JUMP, 1x SLOW.
>>>
>>> Available rates are checked and refilled on every stats window update.
>>=20
>> Very cool!
>>=20
>>> With this approach, we finally get a very small delta in throughput when
>>> comparing setting the optimal data rate as a fixed rate vs normal rate
>>> control operation.
>>=20
>> Can you quantify this "very small delta"? Would love to see some
>> benchmark data :)
> Based on a quick test it seems to be around 5% (sometimes less) at VHT80
> MCS6 nss=3D4 with ~350 mbit/s TCP throughput. I guess I might be able to
> bring that down even further, once I optimize it some more.

Cool! And pre-patch?

-Toke