Re: [HACKERS] performance test

Bruce Momjian wrote:
> 
> > You've seen disadvantages of our unperfect buffer manager -:)
> > When server need in buffer for new data comming and there is
> > no unused buffers in pool (i.e. - all buffers filled with new
> > data and marked as dirty), server gets some dirty buffer,
> > writes it AND FSYNC file. So, server does many fsyncs
> > even with BEGIN/END while should do _one_ fsync at COMMIT.
> >
> > Having this problem fixed you wouldn't had so big difference
> > between -F and BEGIN/END-no-F.
> 
> Is this idea still reasonable?
> 
> ---------------------------------------------------------------------------
> 
> Here is an archive of the pg_log discussion.

Thanks for archive, Bruce.
But delayed fsync is another idea.
bufmgr has to be fixed anyway.

Vadim