Issue metadata
Sign in to add a comment
|
Two-digit regression in octane, kraken and jetstream Javascript test across all boards since 2017/3/11. |
||||||||||||||||||||||||
Issue descriptionCrosbolt dashboard (https://cros-goldeneye.corp.google.com/console/listCrosbolt) showed more than 10% regression in octane/Total/Score on all board, but speedometer score is 10% better. Investigation. Here is an example report from Samus: 10.5% regression in octane/Total/Score on cros-samus at 30350000935600001:30370000935700000 Performance dashboard identified a 10.5% regression in octane/Total/Score on cros-samus at revision range 30350000935600001:30370000935700000. Graph: https://chromeperf.appspot.com/report?masters=ChromeOSPerf&bots=cros-samus&tests=octane%2FTotal%2FScore&checked=Score%2CScore_ref%2Cref&rev=30370000935700000
,
Mar 20 2017
,
Mar 20 2017
,
Mar 20 2017
,
Mar 21 2017
Bisect result: from fb4e0e4554bf71f28e23c64c33b3f0ebbfcb8501 to 9018c99b7f73801b958f5079f68a8fda1519f480 (59.0.3035.0 => 59.0.3037.0) first bad commit: [5e921f820bc0a86566e25b279019ba965535230e] Update V8 to version 5.9.34.2 (Switch on I+TF).
,
Mar 21 2017
V8 team, can you take a look at the regression? It affects all ChromeOS machines.
,
Mar 21 2017
Update. All platforms affected https://chromeperf.appspot.com/report?sid=94cd7cfcf7b385cd52d892e8e60b9995c20aa6f1a432c858bfbb93b4189d3002
,
Mar 21 2017
Yes, this is expected and also was announced some time in advance. We have started to ship V8's new compiler pipeline based on an Interpreter and a next-gen optimizing compiler in M59. See https://v8project.blogspot.de/2017/02/help-us-test-future-of-v8.html as a reference with a little bit more background. The ChromeOS release management is AFAIK aware of it and didn't raise concerns. Keep in mind that the new compiler pipeline will give Chrome wins in more "real-worldly" workloads. See https://v8project.blogspot.de/2016/12/how-v8-measures-real-world-performance.html for more information. If you have concerns or remarks, please get in touch with me.
,
Mar 21 2017
Thank you for your explanation. https://v8project.blogspot.tw/2016/12/how-v8-measures-real-world-performance.html answers the situation I observed: speedometer improved by 10% while other microbenchmarks: octane, kraken and jetstream regressed by 10%. For later performance regression monitoring, shall we just focus on speedometer for JS part?
,
Mar 24 2017
Re #9: Speedometer also tests heavily Blink/the rest of the browser. I would suggest using the new scores (for Octane, Sunpspider, Speedometer, ...) as the new baseline. Regressions from now on shouldn't happen or at least we should be actively be aware of.
,
Apr 5 2017
Issue 708309 has been merged into this issue.
,
Apr 6 2017
|
|||||||||||||||||||||||||
►
Sign in to add a comment |
|||||||||||||||||||||||||
Comment 1 by deanliao@chromium.org
, Mar 20 2017