On Mon, Jan 09, 2023 at 02:40:27PM +0100, Petr Mladek wrote: > Why are try hardly comparable? > > 1. The speed depends on the number of loaded modules > and number of symbols. It highly depends on the configuration > that was used to build the kernel. > > 2. The test runs only once. As a result it is hard to judge > how big is the noise. > > 3. The noise might depend on the size and state of CPU caches. > > > I personally vote for removing this selftest! Even so, just as with testing a filesystem with different types of configurations, at least testing a few configs helps and it's what we do. Then, if anyone ever wanted to try to increase performance on symbol lookup today they have no easy way to measure things. How would they go about comparing things performance without this selftest? This selftests helps generically with that *and* helps peg on to it any sanity checks you may wish to add to those APIs which we just don't want to do upstream. That was the rationale behind it, just as with any other selftest. However, if measuring the time is not possible that's separate topic. But to say that measuring time on some config is not valuable I think is not a fair statement. Yes, the noise things are good points, but the test can be enhanced for that too. And any patch which anyone in the future would try to propose for new enhancements would likely be looking at obvious gains beyond any type of noise. Luis
