New issue
Advanced search Search tips
Note: Color blocks (like █ or █) mean that a user may not be available. Tooltip shows the reason.

Issue 890398 link

Starred by 1 user
Status: WontFix
Owner:
sonnyrao@chromium.org
Closed: Oct 5
Cc:
dtor@google.com
█ ravisadineni@chromium.org
drinkcat@chromium.org
evgreen@chromium.org
█ adityakali@chromium.org
groeck@chromium.org
█ djkurtz@chromium.org
tbroch@chromium.org
█ sawlani@chromium.org
rkolchmeyer@chromium.org
Components:
EstimatedDays: ----
NextAction: ----
OS: Chrome
Pri: 2
Type: Feature



Sign in to add a comment

Do we want PELT_UTIL_HALFLIFE_32, PELT_UTIL_HALFLIFE_16, or PELT_UTIL_HALFLIFE_8?

Project Member Reported by diand...@chromium.org, Sep 28 
As Evan pointed out in <https://chromium-review.googlesource.com/c/chromiumos/third_party/kernel/+/1247067>, we have a choice to make.  This choice was introduced when we merged over the Android EAS patches.

I _believe_ that this choice affects all boards, not just those running EAS.  Someone can correct me if I'm wrong.  

---

Here's the description from the KConfig

---

choice
	prompt "Utilization's PELT half-Life"
	default PELT_UTIL_HALFLIFE_32
	help
	  Allows choosing one of the possible values for the PELT half-life to
	  be used for the update of the utilization of tasks and CPUs.
	  The half-life is the amount of [ms] required by the PELT signal to
	  build up to 50% utilization. The higher the half-life the longer it
	  takes for a task to be represented as a big one.

	  If not sure, use the default of 32 ms.

config PELT_UTIL_HALFLIFE_32
	bool "32 ms, default for server"

config PELT_UTIL_HALFLIFE_16
	bool "16 ms, suggested for interactive workloads"
	help
	  Use 16ms as PELT half-life value. This will increase the ramp-up and
	  decay of utlization and load twice as fast as for the default
	  configuration using 32ms.

config PELT_UTIL_HALFLIFE_8
	bool "8 ms, very fast"
	help
	  Use 8ms as PELT half-life value. This will increase the ramp-up and
	  decay of utlization and load four time as fast as for the default
	  configuration using 32ms.

endchoice

---

...right now we're getting the default (32).  Nearly all Chrome OS devices are about interactivity over server performance.  Should we move to 16?

Comment 1 by sonny...@google.com, Oct 2 

Looks like the choice comes from commit 7ba0536d17ea701f26458a3d9fd2ee58ae98b3a9
    FROMLIST: sched/fair: add support to tune PELT ramp/decay timings

According to patchwork it doesn't look like it's going to land upstream as-is:
https://patchwork.kernel.org/patch/10331795/

I also looked at Android 4.14 and they still have this patch.  I didn't see anything in Android's public repos about later kernels.  I didn't see any defconfigs either.

I don't know much about this, but I do know that one of the reasons Android folks used WALT instead of PELT was because PELT was too slow to adapt to interactive workloads, I'm inclined to believe PELT_UTIL_HALFLIFE_16 is probably the right choice, but do we know if we're going to keep this patch in 4.19? 

Ideally I should get a 4.14 arm system up and do some benchmarking -- it's on my todo list.

Comment 3 by groeck@google.com, Oct 3 

This is part of the EAS code. It is in branch experimental/eas-dev in the Android repository at https://android.googlesource.com/kernel/common. As of the EAS port to 4.18, it is still there.

Presumably this means we'll get it with EAS, upstream or not.

Comment 4 by sonnyrao@chromium.org, Oct 3 

re #3 thanks for the pointer to experimental/eas-dev.  I'm pretty sure we'll want HALFLIFE_16 -- otherwise not sure why they'd keep this patch.  I will try to get my system up and running and see if I can measure any difference.

Comment 5 by groeck@chromium.org, Oct 5 

The patch in question is gone in the EAS port to 4.19. The only option left is a half life of 32, which is hardcoded.

If there is reason t believe that we need this option (and the associated patch) in chromeos-4.19 even though it has been dropped upstream, please let me know.

Comment 6 by sonnyrao@chromium.org, Oct 5

Status: WontFix (was: Untriaged)
ok if it's dropped in 4.19 (not surprising since patchwork says "superseded") then I'm going to close this
►

Sign in to add a comment