Re: pgbench bug / limitation

David Rowley <dgrowleyml(at)gmail(dot)com> writes:
> On Tue, 2 Jun 2020 at 23:27, Fabien COELHO <coelho(at)cri(dot)ensmp(dot)fr> wrote:
>> Attached is a blind attempt at working around the issue based on what you
>> show below, that I cannot test.

> I suppose we might be able to do something like that. It does expose
> us to the implementation of Microsoft's fd_set struct, but surely the
> shape of that must be pretty much fixed since if it were to be changed
> then software that's already compiled would break.

More to the point, this does nothing to get around the O(N^2) cost of
adding N FDs to the fd_set with Microsoft's implementation; so I do not
think it's really going to fix the problem for people who want to use
lots of connections on Windows.

The idea that I vaguely had was to build our own array of socket FDs
(bypassing the unnecessary de-duplication logic in FD_SET) and then call
WaitForMultipleObjects() or similar directly. This would of course
be quite Windows-specific, which'd be annoying, but it seems like
getting out of using select() on Windows wouldn't be a bad thing.
Trying to make the same code work with two basically different models
of what a fd_set is seems like a recipe for pain. This would also get
us out from under the hack of redefining FD_SETSIZE.

Alternatively, if we're going to get in bed with the internals of
Windows' fd_set implementation anyway, we could implement our own
version of FD_SET that doesn't do the de-duplication.

regards, tom lane