-
Notifications
You must be signed in to change notification settings - Fork 26.8k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Regression in diff-min test flutter_gallery__memory_nav #69411
Comments
Or could be #69307, maybe? This is a test change. |
This test is using adb to measure memory so its not caused by something like only looking at one isolate's memory |
It looks this is the difference between memory at the end and memory at the start. If we use less memory at the start (due to not spinning up an isolate) then this would increase. This seems to be confirmed by https://flutter-flutter-perf.skia.org/e/?queries=sub_result%3Dstart-median%26sub_result%3Dstart-min%26test%3Dflutter_gallery__memory_nav |
This is actually a positive change |
Leaving open for future discussion |
FYI @liyuqian , this benchmark can be pretty confusing if we decrease start up memory. Though might be worth talking about as a perf improvement, if I'm reading this correctly we're using 10 MB less memory at startup |
Agree it's confusing as the metric isn't monotone when the performance improves. We've had similar situations before with missed frames metrics which triggered us to replace them with 90th, 99th percentiles (#18727). I'm not sure what's a good replacement though. One question is: if we reduce 10 MB memory at startup, why didn't we also reduce 10 MB memory in the end? Is it that we just defer the memory allocation to a later point? I believe the original intention of measuring the startup memory is to measure the non-Flutter memory (those overhead memory by unrelated Android/iOS stuff) and make our metrics immune to those OS changes. If the startup memory actually includes something that Flutter allocates, maybe we should measure:
We then calculate |
Because it was a temporary memory allocation from the isolate used to parse the asset manifest |
If it's temp, shouldn't it be GCed by the end of the test run? |
It was: The end memory stayed the same, the start memory decreased. So the diff did increase, but none of the component measurements got worse |
I see. So previously the 10MB gets GCed (a -10MB allocation) between start and end, while now that 10MB does not exist anymore. If there's a |
See https://flutter-flutter-perf.skia.org/e/?begin=1603897001&end=1604032912&keys=Xb89e35c986a650a0bf9142de23b3f1bb&requestType=0&xbaroffset=22270
Looks like commit #69264, relanding the staging of web tests.
cc @jonahwilliams, author of the reland commit.
The text was updated successfully, but these errors were encountered: