Re: [PATCH] sched/fair: select waker's cpu for wakee on sync wakeup

[Peng Wang]

On 25/08/2022 17:09, , Mel Gorman wrote:
> On Thu, Aug 25, 2022 at 02:45:05PM +0800, Peng Wang wrote:
> > On 24/08/2022 16:46, , Mel Gorman wrote:
> > > On Wed, Aug 24, 2022 at 12:37:50PM +0800, Peng Wang wrote:
> > > > On sync wakeup, waker is about to sleep, and if it is the only
> > > > running task, wakee can get warm data on waker's cpu.
> > > >
> > > > Unixbench, schbench, and hackbench are tested on
> > > > Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz (192 logic CPUs)
> > > >
> > > > Unixbench get +20.7% improvement with full threads mainly
> > > > because of the pipe-based context switch and fork test.
> > > >
> > > > No obvious impact on schbench.
> > > >
> > > > This change harms hackbench with lower concurrency, while gets improvement
> > > > when concurrency increases.
> > >
> > > Note that historically patches in this direction have been hazardous because
> > > it makes a key assumption "sync wakers always go to sleep in the near future"
> > > when the sync hint is not that reliable. Networking from a brief glance
> > > still uses sync wakeups where wakers could have a 1:N relationship between
> > > work producers and work consumers that would then stack multiple tasks on
> > > one CPU for multiple consumers. The workloads mentioned in the changelog
> > > are mostly strictly-synchronous wakeups (i.e. the waker definitely goes
> > > to sleep almost immediately) and benefit from this sort of patch but it's
> > > not necessarily a universal benefit.
> >
> > Hi, Mel
> >
> > Thanks for your clarification.
> >
> > Besides these benchmarks, I also find a similar strictly-synchronous wakeup
> > case [1].
> >
> > [1]https://www.mail-archive.com/linux-kernel@xxxxxxxxxxxxxxx/msg1478754.html
>
> Yep, but it falls under the same heading, sometimes the caller knows it's
> a strict sync wakeup but not always.
>
> > > Note that most of these hazards occurred *LONG* before I was paying much
> > > attention to how the scheduler behaved so I cannot state "sync is still
> > > unreliable" with absolute certainty. However, long ago there was logic
> > > that tried to track the accuracy of the sync hint that was ultimately
> > > abandoned by commit e12f31d3e5d3 ("sched: Remove avg_overlap"). AFAIK,
> > > the sync hint is still not 100% reliable and while stacking sync works
> > > for some workloads, it's likely to be a regression magnet for network
> > > intensive workloads or client/server workloads like databases where
> > > "synchronous wakeups are not always synchronous".
> > Yes, you are right. Perhaps in such situation, a strong contract from user
> > is a better alternative than struggling with the weak hint in kernel.
>
> Even the kernel doesn't always know if a wakeup is really sync or not
> because it lacks valuable context and the number of tasks on the runqueue is
> insufficient if there are multiple wakeups in quick succession. At best,
> there could be two WF_SYNC hints and hope every caller gets it right
> (hint, they won't because even if it's right once, cargo cult copying
> will eventually get it wrong and there is an API explosion issue such as
> wake_up_interruptible_*). A user hint would be tricky. Core libraries
> couldn't use it because it has no idea if the linked application wants
> a strictly sync wakeup or not, a core library couldn't tell given just
> a pthread_mute_t for example. Even if it was true at one point in time,

OK, I get it now, thanks!

If we passed more information dealing with pthread_mute_t, it would
bring too much changes through user core libraries to this kernel
scheduling decision.

And the current weak sync-wakeup hint can only bring us a candidate
in the same LLC cache domain at most.

> it might not be true later if the application design changed leading to
> application bugs being blamed on the kernel for poor placement decisions.