Re: Question Regarding isolcpus

[Waiman Long <longman@redhat.com>]

On 10/12/23 15:23, Joseph Salisbury wrote:
>
>
> On 10/12/23 15:10, Waiman Long wrote:
>> On 10/12/23 13:27, Joseph Salisbury wrote:
>>>
>>>
>>> On 9/28/23 04:39, Sebastian Andrzej Siewior wrote:
>>>> On 2023-09-26 12:45:14 [-0400], Joseph Salisbury wrote:
>>>>> Hi All,
>>>> Hi,
>>>>
>>>>> I have a question regarding the isolcpus parameter.  I've been 
>>>>> seeing this
>>>>> parameter commonly used. However, in the kernel.org documentation[0],
>>>>> isolcpus is listed as depreciated.
>>>>>
>>>>> Is it the case that isolcpus should not be used at all? I've seen 
>>>>> it used
>>>>> in conjunction with taskset.  However, should we now be telling rt 
>>>>> users to
>>>>> use only cpusets in cgroups?  I see that CPUAffinity can be set in
>>>>> /etc/systemd/system.conf.  Is that the preferred method, so the 
>>>>> process
>>>>> scheduler will automatically migrate processes between the cpusets 
>>>>> in the
>>>>> cgroup cpuset or the list set by CPUAffinity?
>>>> Frederic might know if there is an actual timeline to remove it. The
>>>> suggestions since then is to use cpusets which should be more 
>>>> flexible.
>>>> There was also some work (which went into v6.1 I think) to be able to
>>>> reconfigure the partitions at run-time while isolcpus= is a boot time
>>>> option.
>>>>  From what I remember, you have a default/system cpuset which all 
>>>> tasks
>>>> use by default and then you can add another cpuset for the "isolated"
>>>> CPUs. Based on the partition it can be either the default one or
>>>> isolated [0]. The latter would exclude the CPUs from load balancing
>>>> which is what isolcpus= does.
>>>>
>>>> [0] f28e22441f353 ("cgroup/cpuset: Add a new isolated 
>>>> cpus.partition type")
>>>
>>> This question may be for the cgroups folks.  The kernel.org 
>>> documentation has a WARNING which states: "cgroup2 doesn't yet 
>>> support control of realtime processes and the cpu controller can 
>>> only be enabled when all RT processes are in the root cgroup "[0]. 
>>> Does this mean real-time processes are only supported on cgroupsV1?
>>>
>>> Also, this warning is stated for the "CPU" Controller, but there is 
>>> no mention of this for a "cpuset" controller. Does this imply that 
>>> real-time processes are supported with "cpuset" controllers?
>>
>> Yes, the quoted description applies only to cpu controller. Even for 
>> v1 cpu controller, the realtime support is problematic and there is 
>> no easy solution to that. That is why cgroup v2 doesn't support it.
>>
>> For other controllers, whether the processes are RT or not are 
>> irrelevant. They are equally supported.
>>
>> Cheers,
>> Longman
> Thanks for the feedback, Longman!
>
One further tidbit is the fact that the deadline scheduling policy can 
be used as a replacement of using cgroup v1 cpu controller RT knobs to 
place a limit one how many RT tasks can run on a CPU.

Cheers,
Longman