Made flutter startup faster by allowing initialization to be parallelized

[gaaclarke]

I read through #8489 and tried to get the performance boost in a less invasive way.

We were serializing setup tasks for the shell that shouldn't need to be serialized. This PR lets them fly free as long as possible. It results in a 15% decrease in startup time (~0.05ms).

[chinmaygarde]

Very cool. A couple of nits but looks great.

[chinmaygarde]

Here and elsewhere, please append _future to the ivars or locals that are now futures.

[gaaclarke]

done

[chinmaygarde]

This can just be condensed to one line now: rasterizer_promise.set_value(on_create_rasterizer(...))

[gaaclarke]

done

[chinmaygarde]

Please move this before the platform view creation call. You don't have to wait for the platform view to be created before this can be kicked off as there is no dependency. This can be kicked off eagerly so that platform view creation can be concurrent with rasterizer creation as well.

[gaaclarke]

done

[liyuqian]

Looks great! Do you have some numbers that quantify the improvements of this PR (by probably running some of our start up driver tests locally)? If we don't currently have benchmarks to cover this, we should add some :)

[chinmaygarde]

There is shell_benchmarks but those aren't wired up to any dashboard. There are already benchmarks for startup and shutdown that should show improvements after these changes.

[liyuqian]

Thanks @chinmaygarde ! @gaaclarke : can you please attach comparison numbers of the benchmarks that Chinmay mentioned in the PR description? (You can also find more "startup" tests in https://github.com/flutter/flutter/blob/master/dev/devicelab/manifest.yaml)

[gaaclarke]

@liyuqian

=============== Before =========================
aaclarke-macbookpro2:src aaclarke$ ./out/host_profile/shell_benchmarks
Run on (12 X 2900 MHz CPU s)
2019-09-16 13:48:23
Benchmark                                  Time           CPU Iterations
------------------------------------------------------------------------
BM_ShellInitialization               2733011 ns     103132 ns       1000
BM_ShellShutdown                      718020 ns     281079 ns       2801
BM_ShellInitializationAndShutdown    3829377 ns     369490 ns       1000

=============== After =========================
aaclarke-macbookpro2:src aaclarke$ ./out/host_profile/shell_benchmarks
Run on (12 X 2900 MHz CPU s)
2019-09-16 13:49:41
Benchmark                                  Time           CPU Iterations
------------------------------------------------------------------------
BM_ShellInitialization               2677442 ns      98258 ns       1000
BM_ShellShutdown                      623944 ns     246920 ns       2861
BM_ShellInitializationAndShutdown    3286713 ns     313808 ns       1000

So 0.055569ms faster shell initialization?

[chinmaygarde]

~15% improvement? I'd ship it.

[chinmaygarde]

I think we should add tests that fully complete NotifyCreated on the platform view. That should give us a more usable number about startup costs.

[liyuqian]

Fantastic! Please remember to also put the numbers in the merge description so we can find it in git log. (By default, I think Github won't put your comment in the description.)

[liyuqian]

@gaaclarke : I think @chinmaygarde 's ~15% calculation is based on BM_ShellInitializationAndShutdown, but your ~0.05ms seems to be based on BM_ShellInitialization... In your merged commit message (2620bab), however, they are described as the same thing...