ATH10 firmware question

[Michal Kazior]

On 24 November 2015 at 22:29, Ben Greear <greearb at candelatech.com> wrote:
> On 11/24/2015 10:07 AM, Cedric VONCKEN wrote:
>>
>>         Hi,
>>
>>         I have a simple test platform.
>>         One PC connected to an equipment. This equipment is set in AP
>> mode.
>>         Another PC connected to another equipment. This equipment is set
>> in STA + WDS mode.
>>
>>         Both equipment use the same openwrt Firmware (compat
>> 2015-07-21), I only changed the ath10k firmware (in
>> /lib/firmware/ath10k/...).
>>         Both equipment has the same hardware.
>>         I used a clear channel, and VHT80.
>>         The radio was connected with a coaxial cable and I placed 40 dBm
>> attenuation per Rf chain.
>>         I used the WLN350NX radio card from compex.
>>
>>         First test : ATH10K firmware 10.2.4.70-2 on both equipment
>>                 An iperf from PC connected to the AP to the PC connected
>> to the STA give 919 Mbps.
>>                 An iperf from PC connected to the STA to the PC
>> connected to the AP give 500 Mbps.
>>
>>         Second test : ATHK firmware 10.2.4.70.10-2 on both equipment
>>                 An iperf from PC connected to the AP to the PC connected
>> to the STA give 921 Mbps.
>>                 An iperf from PC connected to the STA to the PC
>> connected to the AP give 441 Mbps.
>>
>>         If I cross the computer I have the same result. I did several
>> time these test and I always have the same result.
>
>
> We see similar.  One thing we notice is that if you actually try to send
> less
> throughput, then you get better overall throughput.
>
> In other words, trying to send 1Gbps UDP frames will give you more poor
> throughput than trying to send 650Mbps (in the upload direction).
>
> I thought it might be a poor interaction regarding backoff in the
> ath10k driver/firmware (see the congestion bins in firmware for why
> this is the case), but even fixing that in firmware didn't improve
> the situation in my testing.

If CPU is the bottleneck on DUT than overcommiting the UDP traffic at
the source may lead the ethernet driver to waste CPU cycles on the
DUT.


Michał