On 11/5/21 10:46 PM, Steev Klimaszewski wrote:
[snip]
Hi,
So IIUC the below logs correctly, you are never hitting boost
frequency (with or without this patch series). Is that correct ?
w.r.t temperature , how are you measuring it? Do you have LMh enabled
or are you using tsens to mitigate cpu temperature ?
Hi,
I was wrong - it does indeed go boost with the patchset applied, it's
just that it doesn't boost up to 2.96GHz very often at all. As noted by
the 0.03% when i ran it while compiling zellij; I reapplied the patches
(and the 6th patch from Lukasz's email) and after boot, 2.96GHz was
showing at 0.39%.
Most tools that read the cpu frequency don't really seem to be well
suited for big.LITTLE, and seem to throw an average of the speed, so
cpufreq-info was the best I have. We're apparently supposed to be using
cpupower these days, but it doesn't seem to know anything about arm64
devices.
Temperature wise, I'm just getting from the sensors, and I am using LMh.
Now, I have to admit, while I've thrown a patch here or there, I'm not
exactly a kernel developer, just enough knowledge to be somewhat
dangerous and know how to backport things. In my mind, and my line of
thinking, I would expect with boost enabled, that the cpu would boost up
to that as often as possible, not require a specific workload to
actually hit it. But then again, I would expect multiple compilation
jobs to be one of the workloads that would?
So I think, the part about never hitting 2.96GHz can be dismissed, and
was simply my lack of knowledge about the cpufreq-info tool's averages.
It does seem however to rarely ever hit 2.96GHz and I would actually
expect it to hit it far more often.
Thank you for the logs and re-testing it with the debug changes.
That's valuable information. I'll work on it today.
Regarding the 'boost' frequency, as you observed, it's hard to
measure it properly. Normally when you use the frequency governor
'schedutil' (which is your case), you need a high utilization workload
to request highest frequency. The task scheduler does this calculation
and then asks for the frequency.
I'll spend more time trying to understand how this Qcom driver and HW
handles it. The patch set itself should not block it.
I'll get back to you soon.
Regards,
Lukasz
