[performance] Rendering is only smooth while you're touching the screen

[Daniel van Vugt]

On some devices (mako at least) it's possible to force double buffering and it will often keep up:

   restart unity8 QML_NO_TOUCH_COMPRESSION=1 MIR_SERVER_NBUFFERS=2

Curiously however it only keeps up smoothly (eg. during a dash scroll) while you're touching the screen. If you lift off or fling the dash then it immediately stutters, quite badly.

So the issue is not that the device can't keep up with double buffering. It seems more like we're not keeping the kernel sufficiently awake and it's clocking down prematurely, as soon as we're not touching it. Certainly adjusting the variables in:
   /sys/devices/system/cpu/cpu0/cpufreq
it is possible to raise the performance and minimum frequency to make double buffering smooth.

The challenge of this bug is to find a way in Mir/QtMir to keep the kernel more awake so it doesn't clock down when we need smooth animations and we're not touching the screen.

Also note that while this bug sounds like bad news for mako, mako is actually the best performing device. Others like arale don't reliably speed up in response to short touches.

[Daniel van Vugt (vanvugt)]

Could be a simple matter of increasing our concurrency. If that's not kept high enough then the delay between context switches likely becomes unacceptably high. And each frame requires at least 4 context switches (USC -> Unity8 -> app -> Unity8 -> USC).

[Daniel van Vugt (vanvugt)]

Further evidence: mirping gets better results while you're touching the screen:

$ mirping /var/run/mir_socket
Connected to server: /var/run/mir_socket
--- Not touching the screen ---
Round-trip time: 0.854571 milliseconds
Round-trip time: 1.373416 milliseconds
Round-trip time: 1.556539 milliseconds
Round-trip time: 1.617580 milliseconds
Round-trip time: 1.434458 milliseconds
Round-trip time: 1.556539 milliseconds
Round-trip time: 1.342896 milliseconds
Round-trip time: 1.373417 milliseconds
Round-trip time: 1.403937 milliseconds
Round-trip time: 1.464978 milliseconds
Round-trip time: 1.526019 milliseconds
--- Now touching the screen a lot ---
Round-trip time: 0.457805 milliseconds
Round-trip time: 0.427285 milliseconds
Round-trip time: 0.579887 milliseconds
Round-trip time: 0.427285 milliseconds
Round-trip time: 0.579887 milliseconds
Round-trip time: 0.518847 milliseconds
Round-trip time: 0.457805 milliseconds
Round-trip time: 0.457805 milliseconds
Round-trip time: 0.518846 milliseconds

[Daniel van Vugt (vanvugt)]

Even Mir's ClientLatency acceptance test has the same behaviour. The latency is almost halved if you're touching the screen while the test runs.

[Daniel van Vugt (vanvugt)]

Three things I have found that solve the stuttering from premature CPU scaling on Android:

  * Don't use nesting. Having a single server architecture seems to keep it smooth.
  * Don't use double buffering (bug 1240909). Staying busy pre-rendering that third buffer keeps clocks higher and smoother.
  * Just wastefully spin a CPU core. Annoyingly, wasting power makes graphics visibly smoother.

Hopefully we can find more options.

[Daniel van Vugt (vanvugt)]

We also have device-specific issues. On mako you will find it spins up secondary CPU cores instantly when you touch the screen. On arale however it waits a few seconds for excessive touching before spinning up the cores, and this keeps frame rates artificially low.

[Thomas Voß (thomas-voss)]

The issue is device-specific and requires integration with SOC-specific infrastructure. We started wiring up the required functionality across the stack, with the tablet being the first real consumer. The changes you are interested in are:

  * http://bazaar.launchpad.net/~unity-system-compositor-team/unity-system-compositor/trunk/revision/280
  * http://bazaar.launchpad.net/~phablet-team/platform-api/trunk/revision/317

We require changes to the Android portions of the platform API, though, with the tablet being the first device that has the required platform api functions implemented on the Android side. The landing for USC can be found here:

  * https://requests.ci-train.ubuntu.com/#/ticket/1097

With the silo installed, device performance is boosted for interactive scenarios, specifically after deep sleep.

[Daniel van Vugt (vanvugt)]

I wonder if this bug might have been solved accidentally in the same way as bug 1388490...

[Daniel van Vugt (vanvugt)]

Possibly fixed now. But we'll need to wait for Mir release 0.26.1 to be able to verify (due to bug 1661128)

[Michał Sawicz (saviq)]

Syncing task from Mir.