The sharded test runner invocations are now passed the previous results
which contains the test timing, which are used to simulate how long each
shard would take to run. The shards are now balanced as evenly as
possible on a test level, taking multiple cores into account.
Sharded tests are now run starting with the slowest test first, such
that extremely long running tests finish as early as possible. This
behavior ensures the cores are saturated and can be padded with fast
tests near the end, rather than waiting for a few slow tests to complete
while the rest of the system is idle.
The algorithm works very well whenever it's able to accurately predict
the time to run shards. In a number of cases, the model doesn't quite
reflect reality and the data, which makes it fairly imperfect but still
reasonably good. I think a second order feedback loop might kick in once
it reorders the tests across shards and the test timing data reflects
the new test timings.
Multitests are no longer always sent to the same shard, since the data
isn't available at the moment, and the change as-is speeds up the test
running considerably.
The front end unit test suites currently ignore the feature as there are
no benefits yet to improving those quick shards.
Upgrade the language version to 3.0.0 so patterns can be used and fix
a mixin not being a mixin.
Fixes: b/291585137
Change-Id: I3cc1b1d96038d5b46e836b091e299097717c226c
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/314081
Reviewed-by: William Hesse <whesse@google.com>
Commit-Queue: Jonas Termansen <sortie@google.com>
Jonas Termansen
f827eb3a78