1255175 - Regression under e10s for FX_REFRESH_DRIVER_SYNC_SCROLL_FRAME_DELAY_MS

Status: RESOLVED FIXED

(Firefox :: General, defect, P1)

Description

[Jim Mathies [:jimm]]

https://gist.github.com/chutten/54f71ccdfa2a6d0b92f9

The non-e10s bars for all three graphs are the same non-e10s numbers. The important thing to note here is that in the child graph, e10s buckets higher than 14msec (60hz refresh times or less) regress compared to non-e10s.

Note this probes data will merge into the main e10s experiment report. chutten is work on on this now.

Comment 1

[Jim Mathies [:jimm]]

The code for this probe landed in bug 1228147.

Comment 2

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

In the 46-noapz analysis [1] the SYNC_SCROLL probe seems to be better with e10s than without e10s. In particular we can see that the e10s values are higher at the "0ms" and "16ms" values, and lower everywhere else. I interpret this as the distribution shifting downwards, bounded by vsync intervals. So for example some of the values in the range 1-15ms shifted down into the 0ms bucket, and some of the values in the 17-31ms shifted down into the 16ms bucket, and so on. We can see from the graph also that the delta between e10s and non-e10s is smaller on the 29ms bucket (which includes the 32ms vsync) than on the 20ms and 24ms buckets, which also supports this theory. As we go further along the graph almost every bucket will contain a vsync interval and so the numbers end up being pretty similar.

I think we can call this fixed, but please reopen if you disagree.

[1] Last three graphs at https://github.com/vitillo/e10s_analyses/blob/70dc53f16ac9cf8045a9e44bf584fae0f0a6e3a1/beta46-noapz/e10s_experiment.ipynb