Date:	Fri, 5 Feb 2016 09:44:43 +0100
From:	Michal Kazior <michal.kazior@...to.com>
To:	Ben Greear <greearb@...delatech.com>
Cc:	"Grumbach, Emmanuel" <emmanuel.grumbach@...el.com>,
	"linux-wireless@...r.kernel.org" <linux-wireless@...r.kernel.org>,
	"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
	Stephen Hemminger <stephen@...workplumber.org>,
	Dave Taht <dave.taht@...il.com>,
	Jonathan Corbet <corbet@....net>
Subject: Re: [RFC v2] iwlwifi: pcie: transmit queue auto-sizing

On 4 February 2016 at 22:14, Ben Greear <greearb@...delatech.com> wrote:
> On 02/04/2016 12:56 PM, Grumbach, Emmanuel wrote:
>> On 02/04/2016 10:46 PM, Ben Greear wrote:
>>> On 02/04/2016 12:16 PM, Emmanuel Grumbach wrote:
>>>>
>>>> As many (all?) WiFi devices, Intel WiFi devices have
>>>> transmit queues which have 256 transmit descriptors
>>>> each and each descriptor corresponds to an MPDU.
>>>> This means that when it is full, the queue contains
>>>> 256 * ~1500 bytes to be transmitted (if we don't have
>>>> A-MSDUs). The purpose of those queues is to have enough
>>>> packets to be ready for transmission so that when the device
>>>> gets an opportunity to transmit (TxOP), it can take as many
>>>> packets as the spec allows and aggregate them into one
>>>> A-MPDU or even several A-MPDUs if we are using bursts.
>>>
>>> I guess this is only really usable if you have exactly one
>>> peer connected (ie, in station mode)?
>>>
>>> Otherwise, you could have one slow peer and one fast one,
>>> and then I suspect this would not work so well?
>>
>>
>> Yes. I guess this one (big) limitation. I guess that what would happen
>> in this case is that the the latency would constantly jitter. But I also

Hmm.. You'd probably need to track per-station packet sojourn time as
well and make it possible to stop/wake queues per station.


>> noticed that I could reduce the transmit queue to 130 descriptors
>> (instead of 256) and still reach maximal throughput because we can
>> refill the queues quickly enough.
>> In iwlwifi, we have plans to have one queue for each peer.
>> This is under development. Not sure when it'll be ready. It also requires
>> firmware change obviously.
>
> Per-peer queues will probably be nice, especially if we can keep the
> buffer bloat manageable.

Per-station queues sound tricky if you consider bufferbloat.

To maximize use of airtime (i.e. txop) you need to send big
aggregates. Since aggregates are per station-tid to maximize
multi-station performance (in AP mode) you'll need to queue a lot of
frames, per each station, depending on the chosen tx rate.

A bursted txop can be as big as 5-10ms. If you consider you want to
queue 5-10ms worth of data for *each* station at any given time you
obviously introduce a lot of lag. If you have 10 stations you might
end up with service period at 10*10ms = 100ms. This gets even worse if
you consider MU-MIMO because you need to do an expensive sounding
procedure before transmitting. So while SU aggregation can probably
still work reasonably well with shorter bursts (1-2ms) MU needs at
least 3ms to get *any* gain when compared to SU (which obviously means
you want more to actually make MU pay off). The rule of thumb is the
longer you wait the bigger capacity you can get.

Apparently there's interest in maximizing throughput but it stands in
direct opposition of keeping the latency down so I've been thinking
how to satisfy both.

The current approach ath10k is taking (patches in review [1][2]) is to
use mac80211 software queues for per-station queuing, exposing queue
state to firmware (it decides where frames should be dequeued from)
and making it possible to stop/wake per-station tx subqueue with fake
netdev queues. I'm starting to think this is not the right way though
because it's inherently hard to control latency and there's a huge
memory overhead associated with the fake netdev queues. Also fq_codel
is a less effective with this kind of setup.

My current thinking is that the entire problem should be solved via
(per-AC) qdiscs, e.g. fq_codel. I guess one could use
limit/target/interval/quantum knobs to tune it for higher latency of
aggregation-oriented Wi-Fi links where long service time (think
100-200ms) is acceptable. However fq_codel is oblivious to how Wi-Fi
works in the first place, i.e. Wi-Fi gets better throughput if you
deliver bursts of packets destined to the same station. Moreover this
gets even more complicated with MU-MIMO where you may want to consider
spatial location (which influences signal quality when grouped) of
each station when you decide which set of stations you're going to
aggregate to in parallel. Since drivers have a finite tx ring this it
is important to deliver bursts that can actually be aggregated
efficiently. This means driver would need to be able to tell qdisc
about per-flow conditions to influence the RR scheme in some way
(assuming a qdiscs even understands flows; do we need a unified way of
talking about flows between qdiscs and drivers?).


[1]: https://www.spinics.net/lists/linux-wireless/msg146187.html
[2]: https://www.spinics.net/lists/linux-wireless/msg146512.html


>>> For reference, ath10k has around 1400 tx descriptors, though
>>> in practice not all are usable, and in stock firmware, I'm guessing
>>> the NIC will never be able to actually fill up it's tx descriptors
>>> and stop traffic.  Instead, it just allows the stack to try to
>>> TX, then drops the frame...
>>
>>
>> 1400 descriptors, ok... but they are not organised in queues?
>> (forgive my ignorance of athX drivers)
>
>
> I think all the details are in the firmware, at least for now.

Yeah. Basically ath10k has a flat set of tx descriptors which are
AC-agnostic. Firmware classifies them internally to per-AC HW queues.


Michał

Powered by blists - more mailing lists