Re: Reducing WaitEventSet syscall churn

On Tue, Jul 14, 2020 at 6:51 PM Thomas Munro <thomas(dot)munro(at)gmail(dot)com> wrote:
> In the meantime, here's a rebase of the more straightforward patches
> in the stack. These are the ones that deal only with fixed sets of
> file descriptors, and they survive check-world on Linux,
> Linux+EXEC_BACKEND (with ASLR disabled) and FreeBSD, and at least
> check on macOS and Windows (my CI recipes need more work to get
> check-world working on those two). There's one user-visible change
> that I'd appreciate feedback on: I propose to drop the FATAL error
> when the postmaster goes away, to make things more consistent. See
> below for more on that.

Here's the effect of patches 0001-0003 on the number of relevant
system calls generate by "make check" on Linux and FreeBSD, according
to strace/truss -f -c:

epoll_create1: 4,825 -> 865
epoll_ctl: 12,454 -> 2,721
epoll_wait: ~45k -> ~45k
close: ~81k -> ~77k

kqueue: 4,618 -> 866
kevent: ~54k -> ~46k
close: ~65k -> ~61k

I pushed those three patches, but will wait for more discussion on the rest.