Re: [PATCH V6 3/6] sched/fair: Fair server interface

[Daniel Bristot de Oliveira]

On 4/11/24 16:43, Peter Zijlstra wrote:
> On Fri, Apr 05, 2024 at 07:28:02PM +0200, Daniel Bristot de Oliveira wrote:
>> Add an interface for fair server setup on debugfs.
>>
>> Each CPU has three files under /debug/sched/fair_server/cpu{ID}:
>>
>>  - runtime: set runtime in ns
>>  - period:  set period in ns
>>  - defer:   on/off for the defer mechanism
>>
>> This then leaves /proc/sys/kernel/sched_rt_{period,runtime}_us to set
>> bounds on admission control.
>>
>> The interface also add the server to the dl bandwidth accounting.
> 
> I suppose most people will want to use it like:
> 
>   for i in /debug/sched/fair_server/cpu*
>   do
> 	  echo $PERIOD > ${i}/period
> 	  ecoh $RUNTIME > ${i}/runtime
>   done
> 
> And I think we agreed to keep this loop in userspace, but memory is
> vague.

correct, we agreed to keep loop in user-space. It is important to have per-cpu
for large systems, like we have at red hat customers (kubernets): each container can
have a different setup... and they do. Like, DPDK people would like to keep some
few us runtime, while other CPUs it is better to keep the default.

> The 'defer' thing is dubious though, I don't suppose anybody would ever
> want to actually change that, other than you while poking around at this
> code, right?

In a setup where all real-time tasks are DL (without fixed-priority tasks (FIFO/RR))
the defer = 0 makes more sense because the bandwidth is reserved anyways, and the
DL server would have a relatively low prio (long period).

Believe it or not, we are getting there in some cases with automation systems :-)

If it does not hurt, I would like keep it... Otherwise, we can think about it in
the future.

-- Daniel