Re: [PATCH v2 6/8] sched/idle: Move busy_cpu accounting to idle callback
From: Aubrey Li
Date: Mon May 17 2021 - 20:59:01 EST
On 5/17/21 8:57 PM, Srikar Dronamraju wrote:
> * Aubrey Li <aubrey.li@xxxxxxxxxxxxxxx> [2021-05-17 20:48:46]:
>
>> On 5/17/21 6:40 PM, Srikar Dronamraju wrote:
>>> * Aubrey Li <aubrey.li@xxxxxxxxxxxxxxx> [2021-05-14 12:11:50]:
>>>
>>>> On 5/13/21 3:31 PM, Srikar Dronamraju wrote:
>>>>> * Aubrey Li <aubrey.li@xxxxxxxxxxxxxxx> [2021-05-12 16:08:24]:
>>>>>> On 5/7/21 12:45 AM, Srikar Dronamraju wrote:
>>>
>>> <snip>
>>>
>>>>>> Also, for those frequent context-switching tasks with very short idle,
>>>>>> it's expensive for scheduler to mark idle/busy every time, that's why
>>>>>> my patch only marks idle every time and marks busy ratelimited in
>>>>>> scheduler tick.
>>>>>>
>>>>>
>>>>> I have tried few tasks with very short idle times and updating nr_busy
>>>>> everytime, doesnt seem to be impacting. Infact, it seems to help in picking
>>>>> the idler-llc more often.
>>>>>
>>>>
>>>> How many CPUs in your LLC?
>>>
>>> I have tried with X86, 48 CPUs, 2 nodes, each having 24 CPUs in LLC
>>> +
>>> POWER10, Multiple CPUs with 4 CPUs in LLC
>>> +
>>> POWER9, Multiple CPUs with 8 CPUs in LLC
>>>
>>>>
>>>> This is a system with 192 CPUs, 4 nodes and each node has 48 CPUs in LLC
>>>> domain.
>>>>
>>>
>>> Okay,
>>>
>>>> It looks like for netperf both TCP and UDP cases have the notable change
>>>> under 2 x overcommit, it may be not interesting though.
>>>>
>>>>
>>>
>>> I believe the extra load on this 24 core LLC could be because we may end up
>>> trying to set the idle-core, even when there is no idle core available.
>>>
>>> If possible, can you please give a try with v3 with the call to
>>> set_next_idle_core commented out?
>>>
>>>
>>
>> v3 seems not be applicable on tip/sched/core 915a2bc3c6b7?
>
> I had applied on top of 2ea46c6fc9452ac100ad907b051d797225847e33
> which was tag: sched-core-2021-04-28
>
> The only conflict you get on today's tip is Gautham's one line patch.
> Gautham's patch replaced 'this' with 'target'.
>
> The 2nd patch does away with that line
>
This is v3. It looks like hackbench gets better. And netperf still has
some notable changes under 2 x overcommit cases.
hackbench (48 tasks per group)
=========
case load baseline(std%) compare%( std%)
process-pipe group-1 1.00 ( 4.51) +1.36 ( 4.26)
process-pipe group-2 1.00 ( 18.73) -9.66 ( 31.15)
process-pipe group-3 1.00 ( 23.67) +8.52 ( 21.13)
process-pipe group-4 1.00 ( 14.65) +17.12 ( 25.23)
process-pipe group-8 1.00 ( 3.11) +16.41 ( 5.94)
process-sockets group-1 1.00 ( 8.83) +1.53 ( 11.93)
process-sockets group-2 1.00 ( 5.32) -15.43 ( 7.17)
process-sockets group-3 1.00 ( 4.79) -4.14 ( 1.90)
process-sockets group-4 1.00 ( 2.39) +4.37 ( 1.31)
process-sockets group-8 1.00 ( 0.38) +4.41 ( 0.05)
threads-pipe group-1 1.00 ( 3.06) -1.57 ( 3.71)
threads-pipe group-2 1.00 ( 17.41) -2.16 ( 15.29)
threads-pipe group-3 1.00 ( 17.94) +19.86 ( 13.24)
threads-pipe group-4 1.00 ( 15.38) +3.71 ( 11.97)
threads-pipe group-8 1.00 ( 2.72) +13.40 ( 8.43)
threads-sockets group-1 1.00 ( 8.51) -2.73 ( 17.48)
threads-sockets group-2 1.00 ( 5.44) -12.04 ( 5.91)
threads-sockets group-3 1.00 ( 4.38) -5.00 ( 1.48)
threads-sockets group-4 1.00 ( 1.08) +4.46 ( 1.15)
threads-sockets group-8 1.00 ( 0.61) +5.12 ( 0.20)
netperf
=======
case load baseline(std%) compare%( std%)
TCP_RR thread-48 1.00 ( 3.79) +4.69 ( 4.03)
TCP_RR thread-96 1.00 ( 4.98) -6.74 ( 3.59)
TCP_RR thread-144 1.00 ( 6.04) -2.36 ( 3.57)
TCP_RR thread-192 1.00 ( 4.97) -0.44 ( 4.89)
TCP_RR thread-384 1.00 ( 19.87) -19.12 ( 28.99)
UDP_RR thread-48 1.00 ( 12.54) -2.73 ( 1.59)
UDP_RR thread-96 1.00 ( 6.51) -6.66 ( 10.42)
UDP_RR thread-144 1.00 ( 45.41) -3.81 ( 31.37)
UDP_RR thread-192 1.00 ( 32.06) +3.07 ( 71.89)
UDP_RR thread-384 1.00 ( 29.57) -21.52 ( 35.50)