Re: [PATCH] sched/fair: Introduce priority load balance for CFS

[Song Zhang]

On 2022/11/16 22:38, Vincent Guittot wrote:
> On Wed, 16 Nov 2022 at 08:37, Song Zhang <zhangsong34@xxxxxxxxxx> wrote:
> > On 2022/11/15 15:18, Vincent Guittot wrote:
> > > On Mon, 14 Nov 2022 at 17:42, Vincent Guittot
> > > <vincent.guittot@xxxxxxxxxx> wrote:
> > > > On Sat, 12 Nov 2022 at 03:51, Song Zhang <zhangsong34@xxxxxxxxxx> wrote:
> > > > > Hi, Vincent
> > > > >
> > > > > On 2022/11/3 17:22, Vincent Guittot wrote:
> > > > > > On Thu, 3 Nov 2022 at 10:20, Song Zhang <zhangsong34@xxxxxxxxxx> wrote:
> > > > > > > On 2022/11/3 16:33, Vincent Guittot wrote:
> > > > > > > > On Thu, 3 Nov 2022 at 04:01, Song Zhang <zhangsong34@xxxxxxxxxx> wrote:
> > > > > > > > > Thanks for your reply!
> > > > > > > > >
> > > > > > > > > On 2022/11/3 2:01, Vincent Guittot wrote:
> > > > > > > > > > On Wed, 2 Nov 2022 at 04:54, Song Zhang <zhangsong34@xxxxxxxxxx> wrote:
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > This really looks like a v3 of
> > > > > > > > > > https://lore.kernel.org/all/20220810015636.3865248-1-zhangsong34@xxxxxxxxxx/
> > > > > > > > > >
> > > > > > > > > > Please keep versioning.
> > > > > > > > > >
> > > > > > > > > > > Add a new sysctl interface:
> > > > > > > > > > > /proc/sys/kernel/sched_prio_load_balance_enabled
> > > > > > > > > >
> > > > > > > > > > We don't want to add more sysctl knobs for the scheduler, we even
> > > > > > > > > > removed some. Knob usually means that you want to fix your use case
> > > > > > > > > > but the solution doesn't make sense for all cases.
> > > > > > > > >
> > > > > > > > > OK, I will remove this knobs later.
> > > > > > > > >
> > > > > > > > > > > 0: default behavior
> > > > > > > > > > > 1: enable priority load balance for CFS
> > > > > > > > > > >
> > > > > > > > > > > For co-location with idle and non-idle tasks, when CFS do load balance,
> > > > > > > > > > > it is reasonable to prefer migrating non-idle tasks and migrating idle
> > > > > > > > > > > tasks lastly. This will reduce the interference by SCHED_IDLE tasks
> > > > > > > > > > > as much as possible.
> > > > > > > > > >
> > > > > > > > > > I don't agree that it's always the best choice to migrate a non-idle task 1st.
> > > > > > > > > >
> > > > > > > > > > CPU0 has 1 non idle task and CPU1 has 1 non idle task and hundreds of
> > > > > > > > > > idle task and there is an imbalance between the 2 CPUS: migrating the
> > > > > > > > > > non idle task from CPU1 to CPU0 is not the best choice
> > > > > > > > >
> > > > > > > > > If the non idle task on CPU1 is running or cache hot, it cannot be
> > > > > > > > > migrated and idle tasks can also be migrated from CPU1 to CPU0. So I
> > > > > > > > > think it does not matter.
> > > > > > > >
> > > > > > > > What I mean is that migrating non idle tasks first is not a universal
> > > > > > > > win and not always what we want.
> > > > > > >
> > > > > > > But migrating online tasks first is mostly a trade-off that
> > > > > > > non-idle(Latency Sensitive) tasks can obtain more CPU time and minimize
> > > > > > > the interference caused by IDLE tasks. I think this makes sense in most
> > > > > > > cases, or you can point out what else I need to think about it ?
> > > > > > >
> > > > > > > Best regards.
> > > > > > >
> > > > > > > > > > > Testcase:
> > > > > > > > > > > - Spawn large number of idle(SCHED_IDLE) tasks occupy CPUs
> > > > > > > > > >
> > > > > > > > > > What do you mean by a large number ?
> > > > > > > > > >
> > > > > > > > > > > - Let non-idle tasks compete with idle tasks for CPU time.
> > > > > > > > > > >
> > > > > > > > > > > Using schbench to test non-idle tasks latency:
> > > > > > > > > > > $ ./schbench -m 1 -t 10 -r 30 -R 200
> > > > > > > > > >
> > > > > > > > > > How many CPUs do you have ?
> > > > > > > > >
> > > > > > > > > OK, some details may not be mentioned.
> > > > > > > > > My virtual machine has 8 CPUs running with a schbench process and 5000
> > > > > > > > > idle tasks. The idle task is a while dead loop process below:
> > > > > > > >
> > > > > > > > How can you care about latency when you start 10 workers on 8 vCPUs
> > > > > > > > with 5000 non idle threads ?
> > > > > > >
> > > > > > > No no no... spawn 5000 idle(SCHED_IDLE) processes not 5000 non-idle
> > > > > > > threads, and with 10 non-idle schbench workers on 8 vCPUs.
> > > > > >
> > > > > > yes spawn 5000 idle tasks but my point remains the same
> > > > >
> > > > > I am so sorry that I have not received your reply for a long time, and I
> > > > > am still waiting for it anxiously. In fact, migrating non-idle tasks 1st
> > > > > works well in most scenarios, so it maybe possible to add a
> > > > > sched_feat(LB_PRIO) to enable or disable that. Finally, I really hope
> > > > > you can give me some better advice.
> > > >
> > > > I have seen that you posted a v4 5 days ago which is on my list to be reviewed.
> > > >
> > > > My concern here remains that selecting non idle task 1st is not always
> > > > the best choices as for example when you have 1 non idle task per cpu
> > > > and thousands of idle tasks moving around. Then regarding your use
> > > > case, the weight of the 5000 idle threads is around twice more than
> > > > the weight of your non idle bench: sum weight of idle threads is 15k
> > > > whereas the weight of your bench is around 6k IIUC how RPS run. This
> > > > also means that the idle threads will take a significant times of the
> > > > system: 5000 / 7000 ticks. I don't understand how you can care about
> > > > latency in such extreme case and I'm interested to get the real use
> > > > case where you can have such situation.
> > > >
> > > > All that to say that idle task remains cfs task with a small but not
> > > > null weight and we should not make them special other than by not
> > > > preempting at wakeup.
> > >
> > > Also, as mentioned for a previous version, a task with nice prio 19
> > > has a weight of 15 so if you replace the 5k idle threads with 1k cfs
> > > w/ nice prio 19 threads, you will face a similar problem. So you can't
> > > really care only on the idle property of a task
> >
> > Well, my original idea was to consider interference between tasks of
> > different priorities when doing CFS load balancing to ensure that
> > non-idle tasks get more CPU scheduler time without changing the native
> > CFS load balancing policy.
> >
> > Consider a simple scenario. Assume that CPU 0 has two non-idle tasks
> > whose weight is 1024 * 2 = 2048, also CPU 0 has 1000 idle tasks whose
> > weight is 1K x 15 = 15K. CPU 1 is idle. Therefore, IDLE load balance is
>
> weight of cfs idle thread is 3, the weight of cfs nice 19 thread is 15

yes, idle weight is 3, thanks for your pointing out.

> > triggered. CPU 1 needs to pull a certain number of tasks from CPU 0. If
> > we do not considerate task priorities and interference between tasks,
> > more than 600 idle tasks on CPU 0 may be migrated to CPU 1. As a result,
> > two non-idle tasks still compete on CPU 0. However, CPU 1 is running
> > with all idle but not non-idle tasks.
> >
> > Let's calculate the percentage of CPU time gained by non-idle tasks in a
> > scheduling period:
> >
> > CPU 1: time_percent(non-idle tasks) = 0
> > CPU 0: time_percent(non-idle tasks) = 2048 * 2 / (2048 + 15000) = 24%
>
> 2 cfs task nice 0 with 1000 cfs idle tasks on 2 CPUs. The weight of
> the system is:
>
> 2*1024 + 1000*3 = 5048 or  2524 per CPU
>
> This means that the cfs nice 0 task should get 1024/(5048) = 20% of
> system time which means 40% of CPUs time.
>
> This also means that the 2 cfs tasks on CPU0 is a valid configuration
> as they will both have their 40% of CPUs

If you increase idle task number to 3000, the cfs nice 0 task only get 
1024 / (2 * 1024 + 3000 * 3) = 9.3% of system time.

But if we can first migrate one cfs nice 0 task to CPU 1, the cfs nice 0 
task maybe execute quickly on CPU 1, then CPU 1 is got to idle and pulls 
more idle tasks from CPU 0, so that the cfs nice 0 task on CPU 0 can 
also be completed more quickly.

> cfs idle threads have a small weight to be negligible compared to
> "normal" threads so they can't normally balance a system by themself
> but by spawning 1000+ cfs idle threads, you make them not negligible
> anymore. That's the root of your problem. A CPU with only cfs idle
> tasks should be seen unbalanced compared to other CPUs with non idle
> tasks and this is what is happening with small/normal number of cfs
> idle threads

If we do not consider putting all low-priority tasks to a cgroup with a 
minimum cpu shares and only set per-task scheduler policy to SCHED_IDLE, 
the weight of a large number of idle tasks cannot be ignored.

> > On the other hand, if we consider the interference between different
> > task priorities, we change the migration policy to firstly migrate an
> > non-idle task on CPU 0 to CPU 1. Migrating idle tasks on CPU 0 maybe
> > interfered with the non-idle task on CPU 1. So we decide to migrate idle
> > tasks on CPU 0 after non-idle tasks on CPU 1 are completed or exited.
> >
> > Now the percentage of the CPU time obtained by the non-idle tasks in a
> > scheduling period is as follows:
> >
> > CPU 1: time_percent(non-idle tasks) = 1024 / 1024 = 100%
> > CPU 0: time_percent(non-idle tasks) = 1024 / (1024 + 15000) = 6.4%
>
> But this is unfair for one cfs nice 0 thread and all cfs idle threads

This unfairness may be short-lived, because as soon as CPU 1 go to idle 
again, CPU 1 immediately pulls more idle tasks from CPU 0 to accelerate 
the running of non-idle tasks on CPU 0.

> > Obviously, if load balance migration tasks prefer migrate non-idle tasks
> > and suppress the interference of idle tasks migration on non-idle tasks,
> > the latency of non-idle tasks can be significantly reduced. Although
> > this will cause some idle tasks imbalance between different CPUs and
> > reduce throughput of idle tasks., I think this strategy is feasible in
> > some real-time business scenarios for latency tasks.
>
> But idle cfs ask remains cfs task and we keep cfs fairness for all threads
>
> Have you tried to :
> - Increase nice priority of the non idle cfs task so the sum of the
> weight of idle tasks remain a small portion of the total weight ?
> - to put your thousands idle tasks in a cgroup and set cpu.idle for
> this cgroup. This should also ensure that the weight of idle threads
> remains negligible compared to others.
>
> I have tried both setup in my local system and I have 1 non idle task per CPU
>
> Regards,
> Vincent

yes I have tried to do them and the results are as expected.

But...as you mentioned above, if all idle tasks are placed in a cgroup 
with the minimum cpu shares or increase nice priority of non-idle tasks, 
the weight of idle tasks is negligible compared with that of non-idle 
tasks, this does not affect the final result.

However, if we try only consider changing the scheduler policy of idle 
tasks to SCHED_IDLE and do not want to modify nice priority of non-idle 
tasks, the weight of idle tasks and the interference on non-idle tasks 
needs to be reconsidered when tasks migration between CPUs.



Best Regards,
Song Zhang

> > > > > Best regards.
> > > > >
> > > > > Song Zhang
> > > .
> .