Hi,
On 7/10/20 2:35 PM, Hans de Goede wrote:
Hi,
On 7/9/20 2:31 PM, Hans de Goede wrote:
Hi,
On 7/7/20 5:43 PM, Rafael J. Wysocki wrote:
Hi Hans,
On Tue, Jul 7, 2020 at 5:32 PM Hans de Goede <hdegoede@xxxxxxxxxx> wrote:
Hi Rafael,
I got a copy of the below bugreport in my mailbox and I
was wondering if this rings any bells.
Are there any relevant power-management / scheduler
changes in 5.7 (vs 5.6) which might be related?
There were a few, but I'm not sure how they can cause thermal issues to happen.
And is there any quick way for the reporter to test this
(by e.g. picking a different scheduler / cpufreq-governor).
Well, as usual, I would recommend looking for possible configuration
differences between the good and bad cases, ie. different default
governors, different cpufreq/cpuidle drivers etc.
Also, if the system is Intel-based, run turbostat to collect CPU PM
information for both the good and bad cases.
Ok I've asked the user to run some tests and provide powertop
and turbostat outputs for both the cold and hot cases.
I will follow-up on this when I know more.
Ok, I've asked the user to gather the necessary info. He has collected
screenshots from both powertop and turbostat under both 5.6 (with the
CPU staying cool) and with 5.7 (where the CPU runs a bit hot) both with
the system fully idle.
The screenshots are attached to the bug as a tarbal:
https://bugzilla.redhat.com/show_bug.cgi?id=1854502
I've never looked at turbostat output before so I do not know what
I'm looking for there. But the powertop output does show a significant
difference.
On the powertop overview page we can see that 5.7 has slightly less
wakeup then 5.6 (probably just normal variation) but despite this,
the "idle stats" page shows that under 5.6 we get 90/91% C7 residency,
where as 5.7 only reaches 85%, that seems like a significant difference
to me. IOW the CPU is in a higher powerstae (consuming more power) for
15% of the time with 5.6, vs 10% with 5.7.
Note that this users system is somewhat customized, he is using
bumblebee with the nvidia binary driver to run games, while
using the i915 gfx for his normal desktop stuff. Doing his
tests no apps are running using the nvidia GPU and in both
cases the nvidia GPU is powered-down (according to the /proc/...
info).
He is also using TLP and some utility to undervolt the GPU. None
of this seems to explain the C7 residency difference though, so
despite his non-standard setup it would be good (IMHO) to still
see if we can try to root-cause that.
The user just reported in bugzilla that the issue is gone with
5.7.8, not sure what fixed it. But now there is no need to look into
this anymore.
Regards,
Hans
