Mateusz Guzik <> wrote:

> You may notice pagezero uses rep. Background page zeroing was removed
> quite some time ago. I see OpenBSD still has it, but it is most likely
> detrimental by now. If it is to be kept, a pagezero variant utilizing
> non-temporal stores still makes sense but a different function should be
> created for on-demand zeroing.

I believe our memory utilization is completely different from other
operating systems.  We are attempting to apply randomization at every
level, and this means we have very different (higher) costs.

Our kernel pool subsystem works differently, different recycling
paradigm.  Our page allocation strategy is not linear.  And finally our
userland malloc stores tracking data in seperate pages and is very
aggressive about munmap, which means we have a much greater amount of
fresh allocation occuring in processes, there is much more zfod mmap()
occuring (we accept the cost of the mmap, but obviously those pages
need to be zero'd by someone).

Everything is different in subtle ways.

Their measurements will probably be very different from ours.

Reply via email to