New issue
Advanced search Search tips
Note: Color blocks (like or ) mean that a user may not be available. Tooltip shows the reason.

Issue 680364 link

Starred by 1 user

Issue metadata

Status: WontFix
Owner:
Last visit > 30 days ago
Closed: Jan 8
Cc:
EstimatedDays: ----
NextAction: ----
OS: Chrome
Pri: 3
Type: Bug-Regression



Sign in to add a comment

10.2% regression in cheets_VellamoTest/Browser_score on cros-veyron_minnie at 29490000907900001:29490000908100000

Project Member Reported by hctsai@chromium.org, Jan 12 2017

Issue description

Performance dashboard identified a 10.2% regression in cheets_VellamoTest/Browser_score on cros-veyron_minnie at revision range 29490000907900001:29490000908100000. Graph: https://chromeperf.appspot.com/report?masters=ARC&bots=cros-veyron_minnie&tests=cheets_VellamoTest%2FBrowser_score&checked=Browser_score%2CBrowser_score_ref%2Cref&rev=29490000908100000

Same regression on other platforms:
https://chromeperf.appspot.com/report?sid=318a01485c8a8ed2fd61b7440df9092fade8eaf899930ed2999633d41f86130c&rev=29510000908700000

Since the regression seems to be happen on ChromeOS side (only ChromeOS uprev between two builds). Currently we don't have good tool for bisect between ChromeOS builds. I would like to write one for current use first and polish it and checkin into the repository in somewhere for sharing later on.
 

Comment 1 by hctsai@chromium.org, Mar 14 2017

The bisection result turns out the regression came from the change:
https://chrome-internal.googlesource.com/chromeos/overlays/project-cheets-private/+/cba99f8682a0a5dcf6f26f3dea0cc4ff961bc6e0

-DIST cheets_arm-target_files-3565428.zip 720882260 SHA256 13cfb4301a848b9f52eb6c7dfbdd94fe5f78687609b1207bdbf581773e5a6289 SHA512 8537088e562bbb679720aae7d432fef74f0e8aea008ea37ed38df1267e42be51c767224abe74921e2b0767eb402ef9e22f57b33d6ece39f4ea516a7bb3703f09 WHIRLPOOL 5b7b5ce1f6ee20c24188084af5fb771c91f0017f8acdc53b8d9da82dd843c559326338e711239bb18da17cc553eb3bd72da21fad5039cc82f253123d5848b420
-DIST cheets_x86-target_files-3565428.zip 846554972 SHA256 afecffd4293b20a7dbf35277f466f5331a24395976ea7441e0f9e289b0a47e1b SHA512 2edeb290b812d5d278377b171b90d0696d365bbae54dd5678c0bec17c0a6d3a37aba3e044266eb74b879764ff9cc2efdd3481121cb8d12d1884eb6b9287cbfbc WHIRLPOOL 266e12eb90815593a53ab4865935b217784b14c4a97e7f6596f47f0bd4ace1bc4cac853a2bab7e04dc53da267a3c2ca6e37aa598afaa6cd0aded2138fb534639
+DIST cheets_arm-target_files-3566735.zip 726002081 SHA256 20c2513cb182f8c4322fb20dbf90b665e9f4cba9ea74b12112cc3f2940ab5314 SHA512 d5bb60f2644574ce98deb4784ac40bff7b989d6ba3b9a357685653c2f0840321b2702142450d10c94c49c0f142e2b9e714efe9f0bdd47202ae440fd6bb9a3b97 WHIRLPOOL 61d3b41a10f7ae063b196ac9d2b74f1b3666ceecdf6dc64d7e6045164fd12ea1c0696fb105ac2ad5f0fe3161090d9e9f72f161d3f347c3322ce93de9036e2b64
+DIST cheets_x86-target_files-3566735.zip 853273200 SHA256 e3924cfb43a1c6a510e910a13355a7dbdcfdf03f2f011a65848bc4fd0783ef8f SHA512 9bf90a859fb899bef81765f499582037a328af2d6f356514bc80959c0939a8dd61e098bc72bdd81e78cbaa2b10235dbf9bb2859a61413439579fa9fb94d1fdb1 WHIRLPOOL daa68e58beae222144bf8773651cf5b077085f2118bc222eabc48d961a619f0213be41f023d638219326c18c4967c0c477d51a7c18ef3accabd98a584391a9bf

The change is to upreve android container from 3565428 to 3566735.

https://android-build-uber.corp.google.com/repo.html?last_bid=3565428&bid=3566735&branch=git_mnc-dr-arc-dev

device/google/cheets2	Y	5bd81fd1	Luis Hector Chavez	Luis Hector Chavez	cheets: Bump up the partition size for the updated WebView
platform/vendor/partner_gms	Y	74d1c6b1	Luis Hector Chavez	Luis Hector Chavez	Uprev WebViewGoogle to 55.0.2883.84
platform/frameworks/av	Y	eaf3bff2	Gurchetan Singh	Gurchetan Singh	ArcCodec: Allow for numFds > 1

Comment 2 by hctsai@chromium.org, Mar 14 2017

Labels: Performance
Owner: lhchavez@chromium.org
Status: Assigned (was: Untriaged)
Hi Luis,

The bug is really old but the Vellemo score didn't come back since then so we are still hunting the root cause. Recently our automated bisection tool for Chrome OS repo is near complete and I decide to come back to old regressions using it for practice runs just like this. 

The Vellamo benchmark runs on Web View thus I think the regression is related to the WebView uprev changes in the list. The problem is that I am not familiar with WebViewGoogle development and have no idea where to track down what have been changed when it is uprevved to 55.0.2883.84 so I am seeking for your help. Please help confirm if the Vellamo score regression could be caused by the WebView uprev or not. If so, is it expected ? Thanks !


Comment 3 by hctsai@chromium.org, Mar 14 2017

Cc: hctsai@chromium.org
Cc: lhchavez@chromium.org
Owner: hctsai@chromium.org
Yeah, it's definitely possible that it was caused by the uprev. The previous version was 53.0.2785.124, so _lots_ of stuff changed :/

https://chromium.googlesource.com/chromium/src/+log/53.0.2785.124..55.0.2883.84?pretty=fuller&n=10000

Comment 5 by hctsai@chromium.org, Mar 15 2017

I just downloaded the APK file and notice there is a libwebviewchromium.so inside, which is produced by the BUILD file in chromium_source/src/android_webview/. Guess doing the bisection only in chromium_source/src/android_webview/ would cause compatibility issues which looks unlikely working.

Does that mean I have to do the bisection against all the changes (20K+ of them) between 53.0.2785.124 and 55.0.2883.84, generate WebViewGoogle APK file accordingly, update the APK file and the whole Chrome binaries and run the Vellamo test ? That's just quite insane. Let me think if there could be better ways to do so.

Labels: OS-Chrome
Labels: -Pri-2 Pri-3
What's left? Can we close this?
Cc: -lhchavez@chromium.org
Status: WontFix (was: Assigned)

Sign in to add a comment