On Mon, 10 Oct 2022, Geliang Tang wrote:

> Geliang Tang <geliang.tang@suse.com> 于2022年10月7日周五 16:59写道：
>>
>> On Thu, Oct 06, 2022 at 05:16:55PM -0700, Mat Martineau wrote:
>>> On Thu, 6 Oct 2022, Geliang Tang wrote:
>>>
>>>> v5:
>>>> - address Mat's comments in v4.
>>>
>>> Hi Geliang -
>>>
>>> Thanks for the v5. I haven't finished looking over all the patches in detail
>>> yet, but two things I do want to reply to right now:
>>>
>>>  * Thanks for explaining in patch 4 that last_snd is still useful for round
>>> robin. I had forgotten about that, and it looked like a "write-only"
>>> variable in the kernel code.
>>>
>>>  * In the meeting today Paolo suggested that a good test for the new
>>> scheduler loop would be to modify simult_flows.sh to use much larger files,
>>> then see if the modified code slowed down any of the simult_flows tests.
>>>
>>> He suggested making the test file 10x larger in simult_flows.sh:
>>>
>>> -     size=$((2 * 2048 * 4096))
>>> +     size=$((2 * 2048 * 4096 * 10))
>>>
>>> Can you compare the test times between the export branch and this series,
>>> with both of them using the larger file size?
>>
>> 10x larger failed in my tests with timeout. So I changed it to 5x
>> larger.
>>
>> Here's the patch:
>>
>> @@ -52,7 +52,7 @@ setup()
>>         sout=$(mktemp)
>>         cout=$(mktemp)
>>         capout=$(mktemp)
>> -       size=$((2 * 2048 * 4096))
>> +       size=$((2 * 2048 * 4096 * 5))
>>
>>         dd if=/dev/zero of=$small bs=4096 count=20 >/dev/null 2>&1
>>         dd if=/dev/zero of=$large bs=4096 count=$((size / 4096)) >/dev/null 2>&1
>> @@ -210,13 +210,13 @@ do_transfer()
>>         fi
>>
>>         echo " [ fail ]"
>> -       echo "client exit code $retc, server $rets" 1>&2
>> -       echo -e "\nnetns ${ns3} socket stat for $port:" 1>&2
>> -       ip netns exec ${ns3} ss -nita 1>&2 -o "sport = :$port"
>> -       echo -e "\nnetns ${ns1} socket stat for $port:" 1>&2
>> -       ip netns exec ${ns1} ss -nita 1>&2 -o "dport = :$port"
>> -       ls -l $sin $cout
>> -       ls -l $cin $sout
>> +       #echo "client exit code $retc, server $rets" 1>&2
>> +       #echo -e "\nnetns ${ns3} socket stat for $port:" 1>&2
>> +       #ip netns exec ${ns3} ss -nita 1>&2 -o "sport = :$port"
>> +       #echo -e "\nnetns ${ns1} socket stat for $port:" 1>&2
>> +       #ip netns exec ${ns1} ss -nita 1>&2 -o "dport = :$port"
>> +       #ls -l $sin $cout
>> +       #ls -l $cin $sout
>>
>>         cat "$capout"
>>         return 1
>>
>> All logs are attached. "5x_10times_export.log" is for the export branch
>> and "5x_10times_refactor_v5.log" is for this series.
>
> I compared the two log data. The test time of this series is indeed
> longer than that of export, but the difference is very small.
>
> I removed the useless information in the two logs and only retained
> the running time and expected time. It looks like this:
>
>> cat 5x_export.log | head
> 36150 max 36005
> 36135 max 36005
> 36087 max 36005
> 36103 max 36005
> 18284 max 18227
> 18349 max 18227
>
> Add all running time and expected time:
>
>> awk '{ sum += $1 }; END { print sum }' 5x_export.log
> 2540920
>> awk '{ sum += $3 }; END { print sum }' 5x_export.log
> 2533820
>
> The ratio of the total running time to the total expected time is
> 1.00280209 (2540920/2533820).
>
> Use the same method to calculate data in 5x_10times_refactor_v5.log:
>
>> cat 5x_refactor_v5.log | head
> 36095 max 36005
> 36168 max 36005
> 36112 max 36005
> 36120 max 36005
> 18235 max 18227
> 18445 max 18227
>
>> awk '{ sum += $1 }; END { print sum }' 5x_refactor_v5.log
> 2546024
>> awk '{ sum += $3 }; END { print sum }' 5x_refactor_v5.log
> 2533820
>
> The ratio is 1.00481644.  It is 0.002 more than export data. I think
> this difference is acceptable.
>

Thanks for taking a look at the data Geliang. To me it seems like the 0.2% 
difference is well within the margin of error, so there doesn't seem to be 
a significant performance regression in your data.

Paolo, how does this performance look to you?


--
Mat Martineau
Intel