async PERF_TEST always returns 0 cycles #2587
Comments
travisdowns
added a commit
to travisdowns/seastar
that referenced
this issue
Dec 17, 2024
Simplify and coroutinize main loop. This reduces the duplicated portion
in the main loop which handles the 4 compile-time cases of
{void returning, size_t returning} x {sync, async}
test case methods. We duplicate only as much as need to keep the main
loop efficient but try to share the remainder. We can also remove a lot
of boilerplate which is no longer needed in C++.
This fixes scylladb#2587 because the underlying bug there was a change in
one of the duplicated sections but not the other.
Performance should be similar after this change: any benchmark that does
substantial work won't see a change. Very short benchmarks are
sensitive to the work in the main loop and this is slightly better in
some cases and slightly worse in others. This was evaluated with
perf_tests_perf benchmark in this series.
We delegate the choice of whether to yield in the benchmark not to the
code under test (in the future returning case).
Fixes scylladb#2587.
travisdowns
added a commit
to travisdowns/seastar
that referenced
this issue
Dec 18, 2024
Simplify and coroutinize main loop. This reduces the duplicated portion
in the main loop which handles the 4 compile-time cases of
{void returning, size_t returning} x {sync, async}
test case methods. We duplicate only as much as need to keep the main
loop efficient but try to share the remainder. We can also remove a lot
of boilerplate which is no longer needed in C++.
This fixes scylladb#2587 because the underlying bug there was a change in
one of the duplicated sections but not the other.
Performance should be similar after this change: any benchmark that does
substantial work won't see a change. Very short benchmarks are
sensitive to the work in the main loop and this is slightly better in
some cases and slightly worse in others. This was evaluated with
perf_tests_perf benchmark in this series.
We delegate the choice of whether to yield in the benchmark not to the
code under test (in the future returning case).
Fixes scylladb#2587.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
If you write a
PERF_TEST(or any other variation likePERF_TEST_F/Cetc) whose testing method is async, the "cycles" column is always zero.The text was updated successfully, but these errors were encountered: