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;
To:     =?UTF-8?Q?Toke_H=c3=b8iland-J=c3=b8rgensen?= <toke@redhat.com>,
        linux-wireless@vger.kernel.org
Cc:     johannes@sipsolutions.net
References: <20210124122812.49929-1-nbd@nbd.name>
 <20210124122812.49929-4-nbd@nbd.name> <87o8hdmdqs.fsf@toke.dk>
From:   Felix Fietkau <nbd@nbd.name>
Subject: Re: [PATCH 4/6] mac80211: minstrel_ht: significantly redesign the
 rate probing strategy
Message-ID: <3ac94bcf-4c76-9c1e-c903-0810d2067b3a@nbd.name>
Date:   Mon, 25 Jan 2021 21:46:41 +0100
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.14; rv:78.0)
 Gecko/20100101 Thunderbird/78.6.1
MIME-Version: 1.0
In-Reply-To: <87o8hdmdqs.fsf@toke.dk>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Precedence: bulk


On 2021-01-25 12:56, Toke Høiland-Jørgensen wrote:
> Felix Fietkau <nbd@nbd.name> writes:
> 
>> The biggest flaw in current minstrel_ht is the fact that it needs way too
>> many probing packets to be able to quickly find the best rate.
>> Depending on the wifi hardware and operating mode, this can significantly
>> 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 changes
>>   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.
> 
> Very cool!
> 
>> 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.
> 
> 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=4 with ~350 mbit/s TCP throughput. I guess I might be able to
bring that down even further, once I optimize it some more.

- Felix