ubuntu-phone team mailing list archive

[Selene Scriven <selene.scriven@xxxxxxxxxxxxx>]

* Matthias Apitz <guru@xxxxxxxxxxx> wrote:
> Yes, exactly like this, with meters in addition.

Although not shown in the picture, I have the phones instrumented 
with meters along the path for primary power and USB, and use it 
to check power consumption on each new build.

For example, here is a summary of one type of krillin power usage 
for 50 builds during a time when a relevant bug was introduced:

  http://toykeeper.net/tmp/phablet/power/krillin-rc-proposed,en-power_usage_display_on-287-336-smooth.png

Toward the right side, it spikes for several builds until we had 
the bug isolated and fixed.  The red columns indicate when it 
thinks there is a new bug, and green is when it thinks something 
was improved.  The blue area is where it expects the result to be 
based on recent measurements.  So, bug found and fixed.  If I 
recall correctly, this particular bug was the reason an OTA 
release got delayed.

Then a few builds later, in r335, it noticed an unusually high 
variation between individual measurements, and marked that build 
for inspection.  For context, here is a more detailed graph for 
r334, when everything was behaving:

  http://toykeeper.net/tmp/phablet/power/krillin.334.display.png

The green section is the part of the measurement it "counts" for 
the test.  Red sections mean USB was plugged in so those values 
aren't relevant.  In this case, it's just letting the phone idle 
with the screen on right after booting.  Five measurements, and 
they're all pretty consistent.

Then in the next build it had one measurement which didn't look 
quite right:

  http://toykeeper.net/tmp/phablet/power/krillin.335.display.png

So it marked that build for inspection, with detailed logs 
available to help identify what happened.

This is how we've been detecting and fixing power consumption 
bugs, making sure each new OTA is the same or better than the 
ones before it.  But that's mostly for userspace bits.  Kernel 
and firmware issues are trickier.

> hat I do not understand is the issue my wife sees from time to 
> time: her device shows 50% or 60% of remaining capacity, for 
> longer time (due to nearly no use of the device), and within 
> minutes the capacity goes to zero and the BQ E4.5 is a brick in 
> her pocket. I understand what you say, Selene, about 
> discrepancies in the layers an error in interpreting the 
> voltage, but I do no see, how can lead this to 50%-to-0% in a 
> few minutes. Have you found something, which explains this?

Yes.  Especially when a device spends a lot of time in standby.  
The daemon which generates that percentage estimate can sometimes 
go a long time between updates...  and when it does update it has 
a tendency to lag.

For example, one day I was testing by manually changing the input 
voltage and recording how the phone responded.  What I found was 
that the kernel's reported voltage lagged behind the actual 
voltage when the actual voltage decreased quickly, but it tracked 
closely when voltage increased.  Additionally, it sometimes took 
a while for the reported percent (and built-in charge graph) to 
catch up.

In this graph comparison, the green line is what the power supply 
voltage was set to, the blue line is what the kernel reported, 
and the rainbow graph is a screenshot of what the phone reports 
to its user.  After manually dropping the voltage to "almost 
empty", it took about 40 minutes for the UI to catch up and it 
did the thing where it went suddenly from like 60% to 0%.

  http://toykeeper.net/tmp/phablet/power/battgraphs.png

Then a bit later (I let it keep measuring), I noticed some other 
odd behavior.  Although it hadn't noticed earlier that the 
voltage went back up, when it finally updated again it changed 
the UI's rainbow graph retroactively:

  http://toykeeper.net/tmp/phablet/power/battgraphs.2.png

After the kernel noticed the increased voltage, the UI took over 
an hour to update.

If it drops suddenly to zero, that usually means the battery 
itself has been dropping for quite a while and the capacity 
estimation software simply took a long time to catch up.

The kernel's reported voltage level isn't perfect, but it's a lot 
closer to reality than the percent shown in the UI.

Of course, there are other factors which make it a bit awkward...  
like the way battery voltage sags under load then recovers later.  
Play an intensive game and the cell voltage might sag to 3.4V...  
but turn the game and screen off and voltage may recover to 3.8V 
within a couple minutes.  So it can be tricky to convert a 
measurable trait (voltage) into an un-measurable trait (percent 
charge remaining).  And the effective capacity isn't a simple 
graph from volts to percent; it changes with the discharge load 
so it's more of a 3-dimensional graph.  And as the cell ages, the 
3D mapping from voltage+amperage to percent changes, so it needs 
recalibration once in a while.


-- Selene