Maintain a single common daemon per test runner, and reuse it without
additional modification wherever possible. Associate each server with
an immutable Description class, and use a new, custom daemon when the
Description corresponding to the test class does not match the
Description corresponding to the method.
Some unscientific benchmarks, running tests after a full build.
On my workstation (HP Z420, 12 cores, 32G RAM, SSD):
buck test --no-results-cache
before: 4:37.88
after: 3:22.36
ChangeIT in Eclipse
before: 26.153s
after: 12.512s
On my laptop (11" MBA, 4 cores, 8G RAM, SSD):
buck test --no-results-cache
before: 15:34.28
after: 11:34.93
ChangeIT in Eclipse
before: 30.903s
after: 19.11s
We still pay a significant startup cost (~6-12s for the server), and
contention during buck test with default -j is an issue (3-5x increase
in time per test). But a 50%/25% reduction in single/aggregate test
time is nothing to sneeze at.
There is still plenty of low-hanging fruit that can take better
advantage of the per-test-class savings, like consolidating more
groups of integration tests together and removing HTTP dependencies.
Change-Id: I23b8d0e9839120a07a3fe1230f19184dc910682c
c02898db13