On 5/18/22 5:48 AM, jaskij wrote:
Hello everyone,
After updating from Hardknott to Kirkstone, we have noticed a massive
performance regression in python3-protobuf.
The regression is caused by the package not building (and thus not
using) the native implementation. What seems to have happened is that
`DISTUTILS_BUILD_ARGS` in the python3-protobuf recipe was not updated to
`SETUPTOOLS_BUILD_ARGS`. This would be a simple fix - just update the
recipe and set the flag, if not for one thing: protobuf in meta-oe is in
version 3.19.4, and python3-protobuf 3.20.0 in meta-python is not able
to link against it.
Internally, I have rolled back python3-protobuf to 3.19.4 and built it
with `SETUPTOOLS_BUILD_ARGS += "--cpp_implementation"`. The difference
is... massive.
For 2000 samples with our test code we get (all times in microseconds):
python3-protobuf 3.20
parse mean: 28713us percentiles: {5: 26623, 25: 27343, 50: 28943,
75: 29663, 95: 30687}
python3-protobuf 3.19
parse mean: 939us percentiles: {5: 756, 25: 944, 50: 950, 75:
956, 95: 1008}
Changing the variable in python3-protobuf is simple, the problem is the
version mismatch. It's either roll back python3-protobuf or update protobuf.
Can you test the updated protobuf scenario ? if that works I would think
thats a preferred aproach.
Jan
--
Jan Dorniak
e-mail:jas...@gmail.com <mailto:e-mail%3ajas...@gmail.com>
-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#97162):
https://lists.openembedded.org/g/openembedded-devel/message/97162
Mute This Topic: https://lists.openembedded.org/mt/91185173/21656
Group Owner: openembedded-devel+ow...@lists.openembedded.org
Unsubscribe: https://lists.openembedded.org/g/openembedded-devel/unsub
[arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
