unity8+Mir draws more current and wakes up 100 times more often than unity8 with surfaceflinger

[Colin Ian King]

I've looked at the CPU, wakeup event and current drawn on a LG Nexus 4 comparing the Mir enabled and non-Mir versions of Unity8 with today's image + updates (at 18:00 UTC, 7th Oct 2013).

I ran a 5 minute idle soak test with the screen set to be un-blanked at full brightness and measured the current drawn at 0.5 second intervals. The Mir enabled version of Unity is drawing on average ~143.7mA where as the non-Mir version is drawing 138.7mA, or around 5mA less current (~3.6% difference)

Measuring CPU load (just Mir)
Mir:
  0.70% user, 0.63% system, 1.33% total CPU usage
Non-Mir:
  0.17% user, 0.07% system, 0.23% total CPU usage

Context Switches:
Mir:
    219.24 context switches/sec
Non-Mir:
    22.32 context switches/sec

poll/epoll/nanosleeps
Mir:
   4.9664/sec
Non-Mir:
   3.1333/sec

Total CPU of system:
Mir:
   0.87%
Non-Mir:
   0.70%

So it seems that the Mir variant of Unity8 is busier and consumes more power than the non-Mir variety when idle.

[Michał Sawicz (saviq)]

You'd need to take surfaceflinger into account when comparing those, as there's no separate display server process now.

[Colin Ian King (colin-king)]

Even factoring out surface flinger, the overall phone is drawing more current now. I will capture the full system activity and report back.

[Colin Ian King (colin-king)]

I've re-measured Mir and non-Mir configurations and double checked my data to include surface flinger. For 300 seconds of complete idle I measured:

Context Switches/sec:
Mir:
   Voluntary: 214.88
   Involuntary: 1.85
   Total: 216.73
Non-Mir + Surface Flinger
   Voluntary: 28.56
   Involuntary: 1.45
   Total: 30.01

..so the Mir variant is context switching over 7 times more than the non-Mir variant.

CPU utilisation:
Mir:
   User: 0.15%
   System: 0.08%
   Total: 0.23%
Non-Mir + Surface Flinger
   User: 0.04%
   System: 0.01%
   Total: 0.05%

..so the Mir variant is consuming ~4.6 times more CPU than the non-Mir variant.

Total System Wakeup Events/sec:
Mir:
  Wakeup Events: 23.57
Non-Mir + Surface Flinger
  Wakeup Events: 0.20

..the Mir variant is creating >100 times more wakeup events than the non-Mir variant

Zero-timeout poll/epoll_wait calls/sec
Mir:
  Zero timeout polls: 2.32
Non-Mir + Surface Flinger
  Zero timeout polls: 1.23

..these are where poll() or epoll_wait() are called with a zero timeout, effectively like doing an immediate peek, which is wasteful because one could be doing a non-blocking read. It may indicate a bug somewhere, as normally one expects poll to be called with a blocking timeout > 0. Anyhow, the Mir variant is worse than the non-Mir variant.

Measured using:
  health-check, cpustat, eventstat from PPA:colin-king/white

Hope that contextualises the bug report.

[Michał Sawicz (saviq)]

Not sure what unity8 itself can do to be better here, as it's almost exactly the same codebase running on both. Have added mir and unity-mir as that's where the investigation needs to go to.

Another component that was replaced by unity-mir (and, in effect, unity8) is ubuntuappmanager, but I doubt that one was eating a lot of power.

[Daniel van Vugt (vanvugt)]

At this early stage a difference of 3.6% is quite good, I think.

A potential fix is in bug 1227739, but see also comment 10 of that.

[Daniel van Vugt (vanvugt)]

Although the difference in wakeups is quite bad.

[Daniel van Vugt (vanvugt)]

Speaking with kgunn the other day he seemed to mention this bug and said the problem was actually a change in the kernel and other packages. Not Mir. Was that the case?

[Daniel van Vugt (vanvugt)]

Do we have any way to retest this any more?

[Kevin DuBois (kdub)]

Not that I know of, the surfaceflinger/unity8 route isn't supported anymore.

[Launchpad Janitor (janitor)]

[Expired for Mir because there has been no activity for 60 days.]

[Launchpad Janitor (janitor)]

[Expired for unity8 (Ubuntu) because there has been no activity for 60 days.]

[Launchpad Janitor (janitor)]

[Expired for The Ubuntu Power Consumption Project because there has been no activity for 60 days.]