Pageset-repeat=5 on no-op CL produces timeTo*Paint regressions |
|||||
Issue descriptionWhile investigating issue 781794 , in https://chromium-review.googlesource.com/c/chromium/src/+/776783 I triggered the perf_bisect bot running tools/perf/run_benchmark try linux system_health.common_desktop --pageset-repeat=5 --story-filter='(.*browse.*social.*twitter.*|.*browse.*news*.cnn.*|.*browse.*news.*hackernews.*)' on a no-op change, i.e. only touch a README, the results.html is here: https://console.developers.google.com/m/cloudstorage/b/chromium-telemetry/o/html-results/results-2017-11-17_09-23-46#r=TOT&s=%25%CE%94avg Activating the storysetRepeats, and looking at the metrics for timeToFirst*Paint, all th0 index runs are producing a > 120 % delta avg. Screenshot below.
,
Nov 20 2017
,
Nov 20 2017
Dave/Simon: any opinion? Is it possible that the bisect hardware has some problem?
,
Nov 20 2017
There could be. Took a look through, found the 2 builds: (with patch): https://build.chromium.org/p/tryserver.chromium.perf/builders/linux_perf_bisect_builder/builds/5424 (without): https://build.chromium.org/p/tryserver.chromium.perf/builders/linux_perf_bisect_builder/builds/5429 Didn't see anything that jumped out, both seemed to be synced to same revision and built from scratch. I'm not an expert on this metric, but looking at the 0th index runs in the results2 provided, it seems like there's just 1 sample that's really off from the rest on the patched version. Specifically, it seems to be from edition.cnn.com. The patched version runs first on the trybot, and looking at the log output it seems to imply that cnn is the very first story run. This might be a problem with the benchmark and the first run being colder than the 2nd?
,
Oct 4
,
Jan 16
(6 days ago)
,
Jan 16
(6 days ago)
|
|||||
►
Sign in to add a comment |
|||||
Comment 1 by drott@chromium.org
, Nov 17 2017