On Wed, Jun 21, 2017 at 11:23 AM, Linus Torvalds <torvalds@linux-foundation.org> wrote: > On Tue, Jun 20, 2017 at 10:22 PM, Andy Lutomirski <luto@kernel.org> wrote: >> There are three performance benefits here: > > Side note: can you post the actual performance numbers, even if only > from some silly test program on just one platform? Things like lmbench > pipe benchmark or something? > > Or maybe you did, and I just missed it. But when talking about > performance, I'd really like to always see some actual numbers.
Here are some timings using KVM:
pingpong between two processes using eventfd: patched: 883ns unpatched: 1046ns (with considerably higher variance)
madvise(MADV_DONTNEED); write to the page; switch CPUs: patched: ~12.5us unpatched: 19us
The latter test is a somewhat contrived example to show off the improved laziness. Current kernels send an IPI on each iteration if the system is otherwise idle.
|