On 2021-06-30, Sven F. <[email protected]> wrote: > On Wed, Jun 30, 2021 at 4:15 AM Stuart Henderson <[email protected]> wrote: >> >> On 2021-06-29, Sven F. <[email protected]> wrote: >> > Dear readers, >> > >> > I probably did something silly again, >> > Could you help with a bit of knowledge around performance ? >> > My openbsd CPU (6.8) is spinning a lot : >> > >> > 0.0%Int 53.1%Spn 25.8%Sys 19.6%Usr 1.4%Idle >> > >> > * Is this bad ? >> > * What kind of basic operation ( like basic shell scripting ) could do >> > that ? ? >> > >> > Thank you, >> > >> >> This means the kernel is spending a lot of time waiting for other CPUs >> to exit locked sections. >> >> First things first, try 6.9, the malloc cache implementation changed >> and that may help. >> >> > > Thank you, > > Can I use ktrace to check where the lock comes from ( can't really put > 6.9 easily to test ) ? > Will try to get 6.9 anyway, but it would be either to know what i need to > test.
If it's due to a userland process then possibly, as might ltrace. To use either of these you will need to already know *what* process is involved, and it is not particularly easy to figure out from the output which calls/functions result in high cpu use. Using dt(4)/btrace and generating a flamegraph gives a view of what's happening in the kernel (it was not enabled by default until after 6.9 so you will need a kernel built with modified config to use it). See https://marc.info/?l=openbsd-bugs&m=158583371404603&w=2 but also see mpi's other posts in the thread. But this is a lot of work to do if it is something that has already been improved in the OS.
