[dpdk-dev] [PATCH v2 3/5] doc: add cycles per packet in testpmd user guide
Ferruh Yigit
ferruh.yigit at intel.com
Mon May 18 11:47:57 CEST 2020
On 5/15/2020 9:33 PM, Dharmik Thakkar wrote:
> Hi Ferruh,
>
>> On May 15, 2020, at 7:15 AM, Ferruh Yigit <ferruh.yigit at intel.com> wrote:
>>
>> On 5/8/2020 11:38 PM, Dharmik Thakkar wrote:
>>> Update documentation for 'show fwd' testpmd runtime function to show
>>> CPU cycles/packet example.
>>>
>>> Signed-off-by: Dharmik Thakkar <dharmik.thakkar at arm.com>
>>> Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli at arm.com>
>>> Reviewed-by: Phil Yang <phil.yang at arm.com>
>>> ---
>>> v2:
>>> - Use shorter lines, up to 100.
>>> ---
>>> doc/guides/testpmd_app_ug/testpmd_funcs.rst | 53 +++++++++++++--------
>>> 1 file changed, 33 insertions(+), 20 deletions(-)
>>>
>>> diff --git a/doc/guides/testpmd_app_ug/testpmd_funcs.rst b/doc/guides/testpmd_app_ug/testpmd_funcs.rst
>>> index a360ecccfd3f..441ed41e3803 100644
>>> --- a/doc/guides/testpmd_app_ug/testpmd_funcs.rst
>>> +++ b/doc/guides/testpmd_app_ug/testpmd_funcs.rst
>>> @@ -351,26 +351,39 @@ Example for the io forwarding engine, with some packet drops on the tx side::
>>>
>>> testpmd> show fwd stats all
>>>
>>> - ------- Forward Stats for RX Port= 0/Queue= 0 -> TX Port= 1/Queue= 0 -------
>>> - RX-packets: 274293770 TX-packets: 274293642 TX-dropped: 128
>>> -
>>> - ------- Forward Stats for RX Port= 1/Queue= 0 -> TX Port= 0/Queue= 0 -------
>>> - RX-packets: 274301850 TX-packets: 274301850 TX-dropped: 0
>>> -
>>> - ---------------------- Forward statistics for port 0 ----------------------
>>> - RX-packets: 274293802 RX-dropped: 0 RX-total: 274293802
>>> - TX-packets: 274301862 TX-dropped: 0 TX-total: 274301862
>>> - ----------------------------------------------------------------------------
>>> -
>>> - ---------------------- Forward statistics for port 1 ----------------------
>>> - RX-packets: 274301894 RX-dropped: 0 RX-total: 274301894
>>> - TX-packets: 274293706 TX-dropped: 128 TX-total: 274293834
>>> - ----------------------------------------------------------------------------
>>> -
>>> - +++++++++++++++ Accumulated forward statistics for all ports+++++++++++++++
>>> - RX-packets: 548595696 RX-dropped: 0 RX-total: 548595696
>>> - TX-packets: 548595568 TX-dropped: 128 TX-total: 548595696
>>> - ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>>> + ------- Forward Stats for RX Port= 0/Queue= 0 -> TX Port= 1/Queue= 0 -------
>>> + RX-packets: 43536504 TX-packets: 43536488 TX-dropped: 0
>>> +
>>> + ------- Forward Stats for RX Port= 1/Queue= 0 -> TX Port= 0/Queue= 0 -------
>>> + RX-packets: 149738504 TX-packets: 149738504 TX-dropped: 0
>>> +
>>> + ------- Forward Stats for RX Port= 1/Queue= 1 -> TX Port= 0/Queue= 1 -------
>>> + RX-packets: 149753052 TX-packets: 149753052 TX-dropped: 0
>>> +
>>> + ---------------------- Forward statistics for port 0 ----------------------
>>> + RX-packets: 43538881 RX-dropped: 72 RX-total: 43538953
>>> + TX-packets: 299491753 TX-dropped: 0 TX-total: 299491753
>>> + ----------------------------------------------------------------------------
>>> +
>>> + ---------------------- Forward statistics for port 1 ----------------------
>>> + RX-packets: 299493085 RX-dropped: 8357 RX-total: 299501442
>>> + TX-packets: 43539683 TX-dropped: 0 TX-total: 43539683
>>> + ----------------------------------------------------------------------------
>>> +
>>> + +++++++++++++++ Accumulated forward statistics for all ports+++++++++++++++
>>> + RX-packets: 343031966 RX-dropped: 8429 RX-total: 343040395
>>> + TX-packets: 343031436 TX-dropped: 0 TX-total: 343031436
>>> + ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>>> +
>>> + CPU cycles/packet=14.28 (total cycles=4899533541 / total RX packets=343031966) at 200 MHz Clock
>>
>> Hi Dharmik,
>>
>> Overall it is OK to show "CPU cycles/packet", but my concern is numbers can be
>> misleading, because for example the numbers I am getting is like below:
>>
>> CPU cycles/packet=4.38 (total cycles=32871036274 / total RX packets=7511734336)
>> at 2100 MHz Clock
>
> CPU cycles/packet shown above is a part of the example. The numbers are as per the statistics shown in the example.
I know but now you are showing a performance number and it is specific to a
platform, my concern it may be misleading and create false expectations for others.
>
>>
>>> +
>>> +.. note::
>>> +
>>> + Measuring CPU cycles/packet requires enabling CONFIG_RTE_TEST_PMD_RECORD_CORE_TX_CYCLES
>>
>> s/TX_CYCLES/CYCLES
>
> Sorry, didn’t understand this comment.
There is no 'CONFIG_RTE_TEST_PMD_RECORD_CORE_TX_CYCLES' config option.
>
>>
>>> + configuration option. On aarch64 platforms, by default, the cycles are counted using
>>> + generic counter which runs at a lower frequency than the CPU clock. To get the cycles/packet
>>> + at CPU clock frequency, please scale the cycles/packet to CPU clock frequency.
>>> + Alternatively, use the PMU based cycle counter.
>>
>> Not sure the "aarch64 platforms" note suits to here, where documents the
>> command. And I can see the platform documentation already updated for this.
>
> This additional comment about aarch64 platforms is to provide complete information to the users with regards to the config option
> and offer a better understanding of the CPU cycles/packet numbers. IMO, without this note, CPU cycles/packet numbers can be misleading.
This is testpmd documentation to describe "show fwd stats" command. We are
adding a note to say with a config option it may also show performance data,
which is OK. But starting to describe the platform differences for this config
option looks to much details for the scope of the document. For testpmd I would
prefer command usage independent from platforms, otherwise the documentation may
become too confusing/complex.
But agree to provide the information for the specific platform, which you are
already doing in other patch in the platform documentation.
>
>>
>> Combining above two comments, what do you think instead of replacing existing
>> stats, add a note saying enabling 'CONFIG_RTE_TEST_PMD_RECORD_CORE_CYCLES'
>> appends "CPU cycles/packet" and give above two samples, like:
>>
>> "
>> Enabling RTE_TEST_PMD_RECORD_CORE_CYCLES appends "CPU cycles/packet" stats, like:
>>
>> CPU cycles/packet=4.38 (total cycles=32871036274 / total RX packets=7511734336)
>> at 2100 MHz Clock
>> OR
>> CPU cycles/packet=14.28 (total cycles=4899533541 / total RX packets=343031966)
>> at 200 MHz Clock
>> "
>
> Yes, this is also an option. IMO, numbers shown within CPU cycles/packet example should be consistent with the forward statistics numbers.
>
This part is just the sample of the "CPU cycles/packet" output, I think it
provides enough context to be understandable without matching Rx packet count in
above part of the stats.
More information about the dev
mailing list