[OpenWrt-Devel] ath9k: fix dynack in IBSS mode

[Koen Vandeputte]

On 28.08.19 19:34, Joe Ayers wrote:
>>> initialized the ackto to max:
>>>
>>> A) avoidance of late-ack state
>>> B) not require wpa_supplicant  -- not in use by our community today
>>> C) Suspect some conditions, e.g. low SNR Neighbors, do not trigger
>>> "late ack" (consistent, with observation of low SNR Neighbors sticking
>>> at max ack_to with my changes )
>>>
>>> flip the algo off/on when new neighbor joins:
>>>
>>> Intended technique to reset ack_to to max.  If ack_to is set to 20km
>>> and then a new adhoc neighbor joins at 30km, this would be a late ack
>>> state, and unable to detect.    My early testing results showed the
>>> algo off/on would restart the ack_to to max and start the process over
>>> with the new neighbor.   I trust I got it right?
>>>
>>> There are 10s to 100s of users testing this bleeding edge change from
>>> nightly builds, and so far, I've not found a failure case.
>>> Although, the findings are showing the cases where static setting has
>>> better throughput.
>>>
>>> Joe AE6XE
>>>
>> Hi Joe,
>>
>> Purely fyi
>>
>> I just pushed dynack improvements to all openwrt branches.
>>
>> I also noticed the issues you addressed above, and these patches fix
>> them for me.
>>
>> Regards,
>>
>> Koen
>>
> Thanks for update.   Updates on performance observations, I've been
> recommending usage of auto settings to the AREDN community as follows:
>
> * best performance gain on Point-to-Point longer distance links (back
> bone links).  I saw ~30% iperf improvement results on a 60km 5GHz link
> -- ack-to floats up under load.   This was about the difference I
> measured on a similar 3GHz  60km link head-to-head comparison between
> AirOS auto distance with TDMA and openwrt static distance with CSMA.
> (3GHz because it takes wifi noise out of the picture.)  I want to do a
> head-to-head comparison again to confirm, but it appears a P2P w/ auto
> setting CSMA in openwrt will compare similar thoughput as AirOS auto
> distance TDMA.

> * good/poor performance for Point-to-Multipoint long distance
> settings, up to 20km  range (cell coverage).   If weak SNR stations, a
> static setting is optimal.  If quality signal, auto works good.
Will try to verify this one
> * Poor performance for short distances, e.g. in the house.   auto
> calculated ack_to settings are several km.  Performance is much poorer
> than a static setting of <1km.
Ack on this one.
Tested on links ranging from ~500m up to 3.5km

static seems to win in terms of performance until the distance goes 
beyond ~6km here


> There seems to be something going on with calculation when 'on the
> bench' testing with short distances.   Maybe a bias needs to be
> applied?

I notice that ack_to never drops below 64 on short distance links. 
(static sets it to 31 on selecting 500m)

I wonder if processing delay/time and thread context switching is coming 
into play here on the slower ar71xx socs

Will check it.

The main focus of this series was to have working links and avoid 
breaking existing ones.

I think the next round will be regarding these performance "issues" 
compared to static

Koen

>
> Joe AE6XE

_______________________________________________
openwrt-devel mailing list
openwrt-devel at lists.openwrt.org
https://lists.openwrt.org/mailman/listinfo/openwrt-devel