MessageLoop start time regression due to code ordering on public bots. |
||
Issue descriptionSee https://chromeperf.appspot.com/report?sid=9df0ec76b0fffc74931cc7bdcf44cd4d540f81bb34ccde3571c2211ec8360ee6&start_rev=531281&end_rev=531574 The culprit is this CL: https://chromium-review.googlesource.com/c/chromium/src/+/874470 Which was reverted in: https://chromium-review.googlesource.com/c/chromium/src/+/886321 The regression is in the ~80ms range for cold start. But... the CL above is actually an improvement. The reason is that it introduces a new way to prefetch the native library on startup. This change makes the prefetch more selective (only .text instead of .text, .rodata and other things), but it also requires code to be ordered in order to kick in. This can be checked by seeing that the prefetch task is <1ms on these bots (and from logcat). The regression we see on the bots are only for the ones running chrome_public_apk, which doesn't use ordering. As a consequence, prefetch doesn't happen. This means that the benefit of prefetching is ~80ms with an unordered native library. Internal bots with code ordered showed a ~40-60ms improvement that disappeared with the revert, see https://chromeperf.appspot.com/report?sid=15c2861109c546f72304a81d85a3692578ce2256e4b012df32602fa8f7b72089&start_rev=1516160166&end_rev=1516881481
,
Jan 25 2018
|
||
►
Sign in to add a comment |
||
Comment 1 by bugdroid1@chromium.org
, Jan 25 2018