Files
palladum-lightning/tests
Rusty Russell 3cd589d2ce bench: why does sending payments take so long under CI?
To run, use:
	VALGRIND=0 TEST_BENCH=1 eatmydata pytest tests/test_connection.py::test_bench

Example of running on my laptop (without --enable-debugbuild, over of 10 runs):

	FAILED tests/test_connection.py::test_bench - assert 47.403406-49.191765(48.343+/-0.48) == 0

With eatmydata:

	FAILED tests/test_connection.py::test_bench - assert 25.867860-27.549878(26.5762+/-0.5) == 0

You can also run perf on l1 once it's running:

	perf record --call-graph dwarf -q -p $(cat /tmp/ltests-*/test_bench_1/lightning-1/lightningd-regtest.pid)

Then ^C after 10 seconds and run "perf report".

Things which stood out:

1. Tracing in db_exec_prepared_v2 (fixed in another PR)

   31.12%     0.04%  lightningd  lightningd            [.] db_exec_prepared_v2
   - 31.08% db_exec_prepared_v2
      + 22.96% db_sqlite3_exec
      + 4.46% trace_span_end
      + 1.77% trace_span_start
      + 1.11% trace_span_tag
      + 0.72% tal_free

2. Logging:

   - 16.03% logv
      - 8.15% maybe_print
         - log_to_files
            + 4.51% __GI__IO_fflush (inlined)
            + 1.97% tal_fmt_
            + 0.51% __GI___strftime_l (inlined)

3. Notification (when nothing is listening) in notify_log:

      - 6.84% maybe_notify_log
         - notify_log
            + 3.37% notify_send
            + 1.75% notify_start
            + 1.71% log_notification_serialize
        0.56% new_log_entry

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2025-05-08 14:01:38 +09:30
..
2025-05-02 13:34:21 -07:00