New submission from Neil Schemenauer <nas-pyt...@arctrix.com>:
I was doing some 'perf' profiling runs of Python. I decided to try running PROFILE_TASK to see what the profile looks like. I was surprised that gc_collect dominated runtime: Children Self Symbol + 93.93% 6.00% [.] _PyEval_EvalFrameDefault + 76.19% 0.12% [.] function_code_fastcall + 70.65% 0.31% [.] _PyMethodDef_RawFastCallKeywords + 63.24% 0.13% [.] _PyCFunction_FastCallKeywords + 58.67% 0.36% [.] _PyEval_EvalCodeWithName + 57.45% 23.84% [.] collect + 52.89% 0.00% [.] gc_collect + 52.10% 0.08% [.] _PyFunction_FastCallDict + 41.99% 0.02% [.] _PyObject_Call_Prepend + 36.37% 0.18% [.] _PyFunction_FastCallKeywords + 20.94% 0.07% [.] _PyObject_FastCallDict + 19.64% 0.00% [.] PyObject_Call + 17.74% 0.11% [.] _PyObject_FastCallKeywords + 12.45% 0.00% [.] slot_tp_call + 12.27% 4.05% [.] dict_traverse + 11.45% 11.04% [.] visit_reachable + 11.18% 10.76% [.] visit_decref + 9.65% 0.11% [.] type_call + 8.80% 0.83% [.] func_traverse + 7.78% 0.08% [.] _PyMethodDescr_FastCallKeywords Part of the problem is that we run full cyclic GC for every test. I.e. cleanup_test_droppings() calls gc.collect(). Maybe we could make these calls conditional on the --pgo flag of regtest. Or, maybe we need to re-evaluate if running the unit test suite is the best way to generate PGO trace data. Based on a tiny bit of further investigation, it looks like gc_collect() is getting called quite a lot of times, in addition to cleanup_test_droppings(). Maybe there is some low-hanging fruit here for optimization. Full GC is pretty expensive. ---------- messages: 336018 nosy: nascheme priority: normal severity: normal status: open title: PROFILE_TASK for PGO build is not a good workload type: performance _______________________________________ Python tracker <rep...@bugs.python.org> <https://bugs.python.org/issue36044> _______________________________________ _______________________________________________ Python-bugs-list mailing list Unsubscribe: https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
