Re: Performance degradation in commit ac1d794

On 2015-12-25 13:28:55 -0500, Tom Lane wrote:
> Hmm. And all those FDs point to the same pipe. I wonder if we're looking
> at contention for some pipe-related data structure inside the kernel.

Sounds fairly likely - and not too surprising. In this scenario we've a
couple 100k registrations/unregistrations to a pretty fundamentally
shared resource (the wait queue for changes to the pipe). Not that
surprising that it becomes a problem.

There's a couple solutions I can think of to that problem:
1) Use epoll()/kqueue, or other similar interfaces that don't require
re-registering fds at every invocation. My guess is that that'd be
desirable for performance anyway.

2) Create a pair of fds between postmaster/backend for each
backend. While obviously increasing the the number of FDs noticeably,
it's interesting for other features as well: If we ever want to do FD
passing from postmaster to existing backends, we're going to need
that anyway.

3) Replace the postmaster_alive_fds socketpair by some other signalling
mechanism. E.g. sending a procsignal to each backend, which sets the
latch and a special flag in the latch structure.

Andres