[Make-wifi-fast] a bit of profiling on the archer
Toke Høiland-Jørgensen
toke at toke.dk
Fri Nov 18 05:26:39 EST 2016
Jesper Dangaard Brouer <brouer at redhat.com> writes:
> On Thu, 17 Nov 2016 20:14:49 -0800 Dave Taht <dave.taht at gmail.com> wrote:
>
>> I have not been profiling much on lower end platforms (it's hard, you
>> can crash a box pretty easily with the wrong options or sample rates).
>
> I'm happy to hear that perf does work on this lower end HW, although
> the disclaimer of sample rates.
>
> Does anyone know if hardware based PMU (Performance Monitor Units)
> exists for these kind of devices?
>
>> While watching the ath10k peak at 150-200mbits, at 99% of cpu in
>> softirq, I spent a bit of time profiling various counters and
>> statistics.
>>
>> for this one (while downloading 12 flows at the same time via flent)
>>
>> perf record -F 99 -e cpu-clock -ag -- sleep 10
>>
>> perf report
>
> Below perf report is not well suite for email, could you instead
> provide output from command below:
>
> perf report --no-children --stdio --call-graph none
>
>>
>> [[31m 67.81%[[m 0.00% ksoftirqd/0 [kernel.kallsyms] [k]
>> run_ksoftirqd
>> |
>> ---run_ksoftirqd
>> |
>> |[[31m--67.61%-- [[m __do_softirq
>> | |
>> | |[[31m--66.80%-- [[m net_rx_action
>> | | |
>> | | |[[31m--41.07%-- [[m ag71xx_poll
>>
>> ...
>
> (Looks like you managed to copy-paste the terminal escape codes for
> colors)
>
>>
>> It appears we're spending 66% of the time in the *ethernet* portion of
>> the path.
>
> Be careful you don't fool yourself. In your output you have the
> "children" mode on, so everything being called "under" net_rx_action is
> summed up. It could be it goes all the way through to the wifi TX
> parts and that is part of the sum. Even the memory allocations gets
> summed into this 66% number.
Yes, actually my guess would be that this is the case. When I was
profiling ath9k I saw this exact behaviour.
-Toke
More information about the Make-wifi-fast
mailing list