1281268 - Measure overhead of GPU process on APZ

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Ryan Hunt [:rhunt] (on leave until early May)]

What would be the overhead of synced ipc between nsBaseWidget and a GPU process, for input events that APZ needs?

Comment 1

[Ryan Hunt [:rhunt] (on leave until early May)]

Attachment: Measure-overhead-of-GPU-process-on-APZ.patch

Created a simple ipc stub, and performed quick measurements of times for certain operations, with and without ipc.

Comment 2

[Ryan Hunt [:rhunt] (on leave until early May)]

Attachment: APZ IPC Test.xlsx

IPC tests were timed as total of:
   1. sending a nop sync message to the GPU process
   2. performing mAPZC->ReceiveInputEvent
NoIPC tests were timed as total of:
   1. performing mAPZC->ReceiveInputEvent

All this happens in nsBaseWidget::DispatchInputEvent

Tests were done individually for MouseEvent, KeyboardEvent, ScrollEvent, with and without IPC enabled.

Summary (seconds):
               Mean             Stddev
ipc mouse      0.0003730906845  0.0001205557166
noipc mouse    0.00001494497903 0.00001190032155
dif            0.0003581457055  0.000108655395
		
ipc keyboard   0.0004262805017  0.0001571841219
noipc keyboard 0.00001429424779 0.00001163543264
dif            0.0004119862539  0.0001455486892
		
ipc scroll     0.0007133422222  0.0002705047813
noipc scroll   0.0001642239336  0.0001003429982
dif            0.0005491182886  0.0001701617831

Comment 3

[Kartikaya Gupta (email:[e-mail])]

So the time spent is increased by an order of magnitude and then some? That's not ideal.

Comment 4

[Kartikaya Gupta (email:[e-mail])]

... but considering that the absolute time increase is still sub-millisecond, we can probably live with it for now. Definitely good to have measured this though, if we find that input event handling time spikes we'll know how much of it to attribute to this IPC overhead.

Comment 5

[Milan Sreckovic [:milan] (needinfo for best results)]

Do we have to worry about the throughput or something like that when there are a lot of events?

Comment 6

[Kartikaya Gupta (email:[e-mail])]

We might, I'm not sure how frequently input events arrive on different platforms. Touch events tend to be throttled to ~1 per vsync interval, so that's not bad. Mousemove events might be more frequent.

Comment 7

[Milan Sreckovic [:milan] (needinfo for best results)]

Is this "just" for scrolling/APZ events, or for any mouse events (e.g., game play)?

Comment 8

[Ryan Hunt [:rhunt] (on leave until early May)]

It's any event on nsBaseWidget::DispatchInputEvent so it was including MouseMoveEvents.

I don't know how fast they come in, but after ten seconds there were a lot of data points.

Comment 9

[Kartikaya Gupta (email:[e-mail])]

And yes, it includes events intended for game play, not just scrolling. We might be able to add a short-circuit path that skips APZ/GPU process if we know there's nothing scrollable in the layer tree. This was something we had thought about doing to reduce event latency on B2G for games but never got around to doing.

Comment 10

[Ryan Hunt [:rhunt] (on leave until early May)]

Now that there is a working version of APZ on the gpu process I thought it would be worth looking at this again.

I did the same idea of test as before, by measuring the time it takes to execute mApz->ReceiveInputEvent (http://searchfox.org/mozilla-central/source/widget/nsBaseWidget.cpp#1251) for different input events and with or without the gpu process. I went to http://news.google.com and scrolled around, moved the mouse, and would type in the search box. The test lasted a minute for with the gpu process, and without.

The results are very similar to before so that's not surprising.

Summary (seconds):
               Mean             Stddev           Data Points
ipc mouse      0.0004690852459  0.0006975921603  2135
noipc mouse    0.0000323615527  0.00001675954085 2705

ipc keyboard   0.0003849004815  0.0001614448421  623
noipc keyboard 0.0000254749499  0.0000150123401  499

ipc wheel      0.0006750419753  0.001069712717   406
noipc wheel    0.00007738832487 0.0000404045465  394

Comment 11

[Ryan Hunt [:rhunt] (on leave until early May)]

Attachment: OOP-APZ Overhead.xlsx

The raw data from the test.

Comment 12

[Kartikaya Gupta (email:[e-mail])]

Thanks for re-running these numbers!