Re: [HACKERS] Slow PITR restore

On Thu, 2007-12-13 at 10:18 -0300, Alvaro Herrera wrote:
> Gregory Stark wrote:
> > "Simon Riggs" <simon(at)2ndquadrant(dot)com> writes:
> >
> > > It's a good idea, but it will require more complex code. I prefer the
> > > simpler solution of using more processes to solve the I/O problem.
> >
> > Huh, I forgot about that idea. Ironically that was what I suggested when
> > Heikki described the problem.
> >
> > I think it's more complex than using posix_fadvise. But it's also more
> > ambitious. It would allow us to use not only the full random access i/o
> > bandwidth but also allow us to use more cpu. In cases where the database fits
> > entirely in ram and we're recovering many many operations modifying the same
> > blocks over and over that might help a lot.
>
> Actually, if you are modifying the same blocks over and over it will
> help *less*, because applying each record needs to occur only after the
> previous records that modify the same block have been applied.
>
> So you have two possibilities: you skip that record and try to apply the
> next one, hoping that that record applies to a block that's not locked,
> (which means you have to remember the skipped record and apply it
> sometime in the future), or you put the process to sleep until the lock
> has been released.

Ah, OK, I can see we're on the same lines of thought there. Just posted
a reply to Heikki about this sort of idea.

--
Simon Riggs
2ndQuadrant http://www.2ndQuadrant.com