Re: 9.4 regression

* Andres Freund (andres(at)2ndquadrant(dot)com) wrote:
> I vote for adapting the patch to additionally zero out the file via
> write(). In your tests that seemed to perform at least as good as the
> old method... It also has the advantage that we can use it a littlebit
> more as a testbed for possibly using it for heap extensions one day.
> We're pretty early in the cycle, so I am not worried about this too much...

I dunno, I'm pretty disappointed that this doesn't actually improve
things. Just following this casually, it looks like it might be some
kind of locking issue in the kernel that's causing it to be slower; or
at least some code path that isn't exercise terribly much and therefore
hasn't been given the love that it should.

Definitely interested in what Ts'o says, but if we can't figure out why
it's slower *without* writing out the zeros, I'd say we punt on this
until Linux and the other OS folks improve the situation.

Thanks,

Stephen