Greetings,
As you can see in the data below, my i4790 box used to default to the
powersave governor despite CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND=y, and
disallowed switching to ondemand. Post 5.8, powersave locks in at the
lowest freq available, and while ondemand becomes the default, it takes
a cross core scheduling performance hit compared to previous powersave.
On a positive note, schedutil, which I used to have cause to avoid,
seems to have improved a lot while I wasn't looking, and now actually
improving cross core performance a wee bit. Seems I should configure
stable 5.[89] kernels to use it as well.
taskset 0xc pipe-test 1
CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND=y defaults to
5.4.74 2.183264 usecs/loop -- avg 2.179480 917.7 KHz powersave, ondemand not allowed
5.5.19 2.184331 usecs/loop -- avg 2.179776 917.5 KHz powersave, ondemand not allowed
5.6.19 2.198156 usecs/loop -- avg 2.204380 907.3 KHz powersave, ondemand not allowed
5.7.19 2.230582 usecs/loop -- avg 2.231981 896.1 KHz powersave, ondemand not allowed
5.8.17 3.203949 usecs/loop -- avg 3.207272 623.6 KHz ondemand
5.8.17 12.840411 usecs/loop -- avg 12.898500 155.1 KHz ondemand -> powersave
5.9.2 3.024442 usecs/loop -- avg 3.040023 657.9 KHz ondemand
5.9.2 12.348519 usecs/loop -- avg 12.299751 162.6 KHz ondemand -> powersave
CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL=y due to db865272d9c4
5.10.gc2dc4c0 2.121057 usecs/loop -- avg 2.121487 942.7 KHz schedutil note: 4.0 GHz perfect!
5.10.gc2dc4c0 2.987989 usecs/loop -- avg 2.988658 669.2 KHz schedutil -> ondemand note: 2.9 GHz not so perfect
5.10.gc2dc4c0 12.447348 usecs/loop -- avg 12.475363 160.3 KHz schedutil -> powersave note: 800 MHz heeelp meee
5.10.gc2dc4c0 2.135548 usecs/loop -- avg 2.137868 935.5 KHz schedutil -> performance note: 4.0 GHz (obviously)
On Sun, 2020-11-01 at 17:23 +0100, Mike Galbraith wrote:
> Greetings,
>
> As you can see in the data below, my i4790 box used to default to the
> powersave governor despite CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND=y, and
> disallowed switching to ondemand.
Ok, my HP lappy running master.today still defaults to powersave, with
CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL=y, still disallows selecting
ondemand as before, and powersave still works. IOW for lappy, it was
business as usual, no change.
Desktop box did, it gained a working ondemand, while its previously
working powersave went broke. Box had schedutil forced upon it, but it
seems perfectly fine with that performance wise.
-Mike
On Monday, November 2, 2020 7:18:41 AM CET Mike Galbraith wrote:
> On Sun, 2020-11-01 at 17:23 +0100, Mike Galbraith wrote:
> > Greetings,
> >
> > As you can see in the data below, my i4790 box used to default to the
> > powersave governor despite CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND=y, and
> > disallowed switching to ondemand.
>
> Ok, my HP lappy running master.today still defaults to powersave, with
> CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL=y, still disallows selecting
> ondemand as before, and powersave still works. IOW for lappy, it was
> business as usual, no change.
OK
> Desktop box did, it gained a working ondemand, while its previously
> working powersave went broke.
Most likely that's because it was handled by intel_pstate in the "active" mode
previously, while it is now handled by it in the "passive" mode (the modes are
described in detail in Documentation/admin-guide/pm/intel_pstate.rst) and the
default governor should be "schedutil".
Which should be slightly better from the functionality perspective.
> Box had schedutil forced upon it, but it
> seems perfectly fine with that performance wise.
Which is the expected outcome (sorry for the confusion).
Cheers!
On Thu, 2020-11-05 at 15:31 +0100, Rafael J. Wysocki wrote:
> On Monday, November 2, 2020 7:18:41 AM CET Mike Galbraith wrote:
>
> > Desktop box did, it gained a working ondemand, while its previously
> > working powersave went broke.
>
> Most likely that's because it was handled by intel_pstate in the "active" mode
> previously, while it is now handled by it in the "passive" mode...
Perhaps the user interface should then nak switching to powersave as it
used to nak switching to ondemand?
-Mike
On Thursday, November 5, 2020 4:08:30 PM CET Mike Galbraith wrote:
> On Thu, 2020-11-05 at 15:31 +0100, Rafael J. Wysocki wrote:
> > On Monday, November 2, 2020 7:18:41 AM CET Mike Galbraith wrote:
> >
> > > Desktop box did, it gained a working ondemand, while its previously
> > > working powersave went broke.
> >
> > Most likely that's because it was handled by intel_pstate in the "active" mode
> > previously, while it is now handled by it in the "passive" mode...
>
> Perhaps the user interface should then nak switching to powersave as it
> used to nak switching to ondemand?
It cannot do that if the powersave governor is configured in.
[Essentially, the problem is that the "powersave" thing advertised by
intel_pstate in the "active" mode is not really the powersave governor,
but that is a mistake made in the past and cannot be undone. Sorry about
that.]
On Thu, 2020-11-05 at 19:02 +0100, Rafael J. Wysocki wrote:
> On Thursday, November 5, 2020 4:08:30 PM CET Mike Galbraith wrote:
> > On Thu, 2020-11-05 at 15:31 +0100, Rafael J. Wysocki wrote:
> > > On Monday, November 2, 2020 7:18:41 AM CET Mike Galbraith wrote:
> > >
> > > > Desktop box did, it gained a working ondemand, while its previously
> > > > working powersave went broke.
> > >
> > > Most likely that's because it was handled by intel_pstate in the "active" mode
> > > previously, while it is now handled by it in the "passive" mode...
> >
> > Perhaps the user interface should then nak switching to powersave as it
> > used to nak switching to ondemand?
>
> It cannot do that if the powersave governor is configured in.
>
> [Essentially, the problem is that the "powersave" thing advertised by
> intel_pstate in the "active" mode is not really the powersave governor,
> but that is a mistake made in the past and cannot be undone. Sorry about
> that.]
Hohum. A little unfortunate, but it probably only affects a few aging
boxen like mine, and I now know better that to ever again do that.