Re: Here are some ath10k performance graphs.

[Ben Greear <greearb@candelatech.com>]

On 03/30/2014 11:27 PM, Ben Greear wrote:
> I finally have a firmware stable enough to do some throughput
> tests (only locked station machine up once doing these!).
>
> I find the download test interesting where the rx link rate goes very low
> after 12 or so stations.
>
> tx link rate is not reported properly yet by ath10k, so ignore that.
>
> http://www.candelatech.com/examples/cap-ath10k-tcp-download-open/
> http://www.candelatech.com/examples/cap-ath10k-tcp-download-wpa2/
> http://www.candelatech.com/examples/cap-ath10k-tcp-upload-open/
> http://www.candelatech.com/examples/cap-ath10k-tcp-upload-wpa2/

Here are some UDP graphs, both with no encryption.  Upload is still
significantly better than download, in this case, possibly because
it is safe to send larger UDP PDUs (I used 9k) from stations.  Sending
large MTUs too stations from wired ports causes too much loss because
loss of a single sub-frame causes entire pkt to be dropped.  Previous
detailed testing shows that is not the only problem, however.

http://www.candelatech.com/examples/cap-ath10k-udp-upload-open/
http://www.candelatech.com/examples/cap-ath10k-udp-download-open/

All of these have a summary chart near the bottom that is generally
more interesting than the individual pie charts, btw.

I can get more throughput with carefully tuned individual connections
for single-station tests, will post some of those when I get time
to do some detailed testing.

Thanks,
Ben

>
> Some notes are near the top of each report, and in general:
>
> The station machine is sending to/from wired Eth1 to/from the
> stations, ie it is both sides of the network.
> The AP is configured as router.  Both systems are running
> our customized 10.1.467-ct firmware, and near the latest ath10k kernel.  Station
> machine is Fedora 17, AP is Fedora 19.  Both are 64-bit, CPU is a mobile core-i7,
> 8GB RAM, SSD hard drives (not that it should matter).
>
> Due to the vagaries of ath10k hardware and our firmware, we get hardware acceleration when
> sending encrypted packets, but we have to decrypt received packets on the CPU.
>
> I am not sure why open TCP throughput is so bad compared to upload.
> I'll run some UDP tests when I get a chance.
>
>
> Thanks,
> Ben
>


-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com