mean_input_event_latency regression from M40 to M41
7 views
Skip to first unread message
Timothy Dresser
Mar 27, 2015, 10:33:39 AM3/27/15
to input-dev
The TouchToScrollUpdateSwap UMA metric is showing a significant regression from M40 to M41 for some devices, so I dug into it a bit in telemetry.
With 10 runs on an S5, I'm seeing:
I took some quick traces, which only showed an 8ms difference in average InputLatency. Does anyone have any thoughts on what could be causing this regression?
I can try to get some traces that show the issue more clearly if necessary.
I can try to get some traces that show the issue more clearly if necessary.
Timothy Dresser
Mar 27, 2015, 10:47:48 AM3/27/15
to input-dev
Correction: this data comes from an S4 (i9500).
Jared Duke
Mar 27, 2015, 11:53:44 AM3/27/15
to Timothy Dresser, input-dev, vmi...@chromium.org, Alexandre Elias, sie...@chromium.org
Looking at the data, there's about a 2ms regression in the mean, but a ~9ms regression in the 90th percentile for the S4.
Alex, Victor: When did we first ship GPU raster (for selected mobile sites)? Was it M41? I believe we blacklisted GPU raster for PowerVR SGX devices, but I think there's an Adreno variant for the S4. If there's any we we could bin the latency results based on GPU+device, that might help. I know we also enabled the GPU process for M41.
Looking at the trace, the GPU thread appears much more saturated in M41. I see a number of "GLES2::CreateStreamTextureCHROMIUM" calls that take ~30-50ms to take complete in M41, but just ~5-10ms in M40.
Timothy Dresser
Mar 27, 2015, 12:01:04 PM3/27/15
to Jared Duke, input-dev, vmi...@chromium.org, Alexandre Elias, sie...@chromium.org
The regression appears to be due to GPU raster.
10 runs, locally, on techcrunch, with M41.
10 runs, locally, on techcrunch, with M41.
With GPU raster set to default: 46.91ms ± 7.15%
With GPU raster disabled: 19.98ms ± 30.12%
Jared Duke
Mar 27, 2015, 12:16:41 PM3/27/15
to Timothy Dresser, input-dev, vmiura, Alexandre Elias, sie...@chromium.org
Thanks for investigating Tim, can you file a bug where we can track further progress?
Alexandre Elias
Mar 27, 2015, 12:19:43 PM3/27/15
to Jared Duke, Timothy Dresser, input-dev, vmiura, Daniel Sievers
As for why it got enabled in M41, I've been slowly increasing the device whitelist release by release and I'm not sure exactly what in M41 might've affected the S5, without knowing the variant it is from chrome://gpu.
Timothy Dresser
Mar 27, 2015, 12:58:54 PM3/27/15
to Alexandre Elias, Jared Duke, input-dev, vmiura, Daniel Sievers
Timothy Dresser
Mar 27, 2015, 2:00:27 PM3/27/15
to Alexandre Elias, Jared Duke, input-dev, vmiura, Daniel Sievers
Sorry, it looks like the data that pointed to GPU raster was just a _really_ bad set of samples. GPU raster has no impact.
I am consistently able to repro the regression from M40 to M41 though. Any other suggestions for potential culprits?
I am consistently able to repro the regression from M40 to M41 though. Any other suggestions for potential culprits?
Victor Miura
Mar 27, 2015, 3:35:53 PM3/27/15
to Timothy Dresser, Alexandre Elias, Jared Duke, input-dev, Daniel Sievers
Could you also try with M42?
Timothy Dresser
Mar 27, 2015, 4:03:05 PM3/27/15
to Victor Miura, Alexandre Elias, Jared Duke, input-dev, Daniel Sievers
I just ran two separate runs of 10 iterations each on M42, and it appears that the regression is still there.
40.76ms ± 12.93%
40.76ms ± 12.93%
37.36ms ± 16.11%
Reply all
Reply to author
Forward
0 new messages