RE: [PATCH 0/2] introduce PM QoS interface

["Morten Brørup" <mb@smartsharesystems.com>]

> From: lihuisong (C) [mailto:lihuisong@huawei.com]
> Sent: Tuesday, 26 March 2024 13.15
> 
> 在 2024/3/26 16:27, Morten Brørup 写道:
> >> From: lihuisong (C) [mailto:lihuisong@huawei.com]
> >> Sent: Tuesday, 26 March 2024 03.12
> >>
> >> 在 2024/3/22 20:35, Morten Brørup 写道:
> >>>> From: lihuisong (C) [mailto:lihuisong@huawei.com]
> >>>> Sent: Friday, 22 March 2024 09.54
> > [...]
> >
> >>>> For the case need PM QoS in DPDK, I think, it is better to set cpu
> >>>> latency to zero to prevent service thread from the deeper the idle
> >> state.
> >>> It would defeat the purpose (i.e. not saving sufficient amounts of
> >> power) if the CPU cannot enter a deeper idle state.
> >> Yes, it is not good for power.
> >> AFAIS, PM QoS is just to decrease the influence for performance.
> >> Anyway, if we set to zero, system can be into Cstates-0 at least.
> >>> Personally, I would think a wake-up latency of up to 10 microseconds
> >> should be fine for must purposes.
> >>> Default Linux timerslack is 50 microseconds, so you could also use
> >> that value.
> >> How much CPU latency is ok. Maybe, we can give the decision to the
> >> application.
> > Yes, the application should decide the acceptable worst-case latency.
> >
> >> Linux will collect all these QoS request and use the minimum latency.
> >> what do you think, Morten?
> > For the example application, you could use a value of 50 microseconds
> and refer to this value also being the default timerslack in Linux.
> There is a description for "/proc/<pid>/timerslack_ns" in Linux document
> [1]
> "
> This file provides the value of the task’s timerslack value in
> nanoseconds.
> This value specifies an amount of time that normal timers may be
> deferred in order to coalesce timers and avoid unnecessary wakeups.
> This allows a task’s interactivity vs power consumption tradeoff to be
> adjusted.
> "
> I cannot understand what the relationship is between the timerslack in
> Linux and cpu latency to wake up.
> It seems that timerslack is just to defer the timer in order to coalesce
> timers and avoid unnecessary wakeups.
> And it has not a lot to do with the CPU latency which is aimed to avoid
> task to enter deeper idle state and satify application request.

Correct. They control two different things.

However, both can cause latency for the application, so my rationale for the relationship was:
If the application accepts X us of latency caused by kernel scheduling delays (caused by timerslack), the application should accept the same amount of latency caused by CPU wake-up latency.

This also means that if you want lower latency than 50 us, you should not only set cpu wake-up latency, you should also set timerslack.

Obviously, if the application is only affected by one of the two, the application only needs to adjust that one of them.

As for the 50 us value, someone in the Linux kernel team decided that 50 us was an acceptable amount of latency for the kernel; we could use the same value, referring to that. Or we could choose some other value, and describe how we came up with our own value. And if necessary, also adjust timerslack accordingly.