Re: [PATCH v3 1/2] sched/fair: Introduce short duration task check

[Honglei Wang]

On 2022/12/13 13:46, Chen Yu wrote:
> On 2022-12-12 at 10:17:35 -0800, Josh Don wrote:
> > > BTW, I've changed the threshold to (sysctl_sched_min_granularity / 8) in my next
> > > version, as this is the value that fit my previous test case and also not to break
> > > the case Josh mentioned.
> >
> > Do you mean a hardcoded value of some number of micros, or literally
> > sched_min_granularity / 8?
> The latter. According to the test, the average task duration when system
> is under heavy load:
> 6 ~ 9 us for netperf
> 7 ~ 70 us for hackbench
> 7 ~ 8 us for tbench
> 13 ~ 20 ms for schbench
> Overall the duration of the micros are quite small(except for schbench).
> The default sysctl_sched_min_granularity is 750 us in kernel if no user
> has changed it. Then 750 / 8 = 93 us, which is close to what you suggested(100us).
> On the other hand, if someone changes sysctl_sched_min_granularity,
> then '8' can be viewed as log2(256). That is, if there are 256 CPUs online,
> and the sysctl_sched_min_granularity is changed to 750 us * log2(256) by
> the user, we can devide the sysctl_sched_min_granularity by 8 in case the
> sysctl_sched_min_granularity is too large.

Hi Yu,

Seems there is a min_t() call in get_update_sysctl_factor(). In most 
cases, we'll get 750 us * (1+log2(8)) = 3000 us in default due to 
sysctl_sched_tunable_scaling is set as '1' default. (Correct me if I 
misunderstand).

For the value in production environment, I've seen 10 ms and 3 ms in 
different place, FYI. Hope this help.

Thanks,
Honglei

> My concern of using hardcoded value is that, this value depends on how fast
> the CPU runs(cpu frequency). The value I measured above is when the
> CPU is running at 1.9Ghz. If a CPU runs faster, a hard code value might not
> be appropriate and could not be tuned.
> > I don't think the latter is necessary, and
> > indeed can lead to weirdness if min_gran is too small or too large. I
> > don't think the concept of what a short duration task is should
> > expand/contract with min_gran.
> The value of sysctl_sched_min_granularity might indicate how long the
> user would like a task to run at least. If the user enlarge this value,
> does it mean the user wants every task in the system to run longer?
>  From this point I found connection between the the definition of short task
> duration and this value. I'm open to changing this value to a fixed one, may
> I have more insights on how this value would be set in production environment?
>
> thanks,
> Chenyu
> > Best,
> > Josh